https://live.cms.zeaware.net/zeaware en-AU https://zeaware.com/cms?v=5.6.1 Wed, 09 Sep 2026 00:00:00 GMT hourly 1 OpenAI GPT-6 Astra now supported in Zeaware Avalon https://live.cms.zeaware.net/zeaware/blog/openai-gpt-astra-supported Wed, 09 Sep 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/openai-gpt-astra-supported

Zeaware Avalon now supports GPT-6 Astra, the latest frontier model released by OpenAI.

This continues Avalon’s approach of supporting a range of leading AI models, allowing organisations to select the most appropriate model for each agent, workflow and use case.

What’s changed with GPT-6 Astra

OpenAI describes GPT-6 Astra as its most intelligent and aligned model to date, with significant improvements across computer use, browsing, software engineering, scientific reasoning and professional work.

Some of the most relevant improvements for enterprise AI include:

  • Stronger reasoning across complex, multi-step tasks
  • Significant improvements in computer and browser use
  • Improved software engineering and codebase understanding
  • Stronger performance across professional knowledge-work tasks
  • Improved ability to work across tools and applications
  • Better adherence to task boundaries and user intent

Astra is particularly interesting for agentic applications because the improvements extend beyond generating better responses. The model is increasingly capable of completing end-to-end workflows involving reasoning, tools, software and external systems.

Built for agent workflows

GPT-6 Astra supports many of the capabilities increasingly important to enterprise agents, including function calling, structured outputs, web and file search, computer use, MCP and Skills.

It also introduces capabilities such as asynchronous tool calling, allowing the model to continue useful work while an application is waiting for another tool to complete.

This direction aligns closely with Avalon’s agent architecture, where the underlying model works alongside tools, skills, enterprise knowledge, memory and governed execution.

How this fits within Zeaware Avalon

GPT-6 Astra is available as a selectable model within Zeaware Avalon and can be used across scenarios including:

  • Enterprise AI agents
  • Tool and skill-enabled workflows
  • Retrieval-based (RAG) solutions
  • Document and content generation
  • Research and analysis
  • Software and technical workflows
  • General-purpose AI interactions

Existing Avalon orchestration and governance patterns remain in place, allowing Astra to be introduced without redesigning the surrounding agent architecture.

This is important for enterprise deployments. The model is one component of an AI solution. Organisations still need appropriate grounding, tools, permissions, security controls, monitoring and governance around what an agent is allowed to do.

A significant step forward for enterprise agents

Perhaps the most interesting aspect of Astra is the continuing shift from models that primarily answer towards models that can reason and act.

For enterprise applications, this potentially means agents capable of completing increasingly substantial pieces of work rather than simply assisting users with individual steps.

The practical opportunity is not to give a model unrestricted autonomy, but to combine these stronger capabilities with well-defined tools, skills, knowledge and governance.

That is the model we are continuing to pursue with Zeaware Avalon.

What to consider

As with any major model release, organisations should evaluate Astra against their actual workloads rather than relying purely on benchmark results.

  • Testing existing agent workflows against Astra
  • Comparing quality with other models available within Avalon
  • Measuring end-to-end task performance rather than individual responses
  • Reviewing latency and token costs
  • Revalidating tool use and structured outputs
  • Applying appropriate controls to higher-autonomy workflows

Astra also represents a substantial increase in cybersecurity capability. OpenAI classifies it at the Critical capability level for cybersecurity under its Preparedness Framework, making appropriate safeguards particularly important for applications involving powerful tools or system access.

Availability

GPT-6 Astra support is now available in Zeaware Avalon.

Organisations using existing OpenAI, Anthropic or other supported models can evaluate Astra through Avalon’s normal model configuration and testing process without changing the broader agent architecture.

]]>

Zeaware Avalon now supports GPT-6 Astra, the latest frontier model released by OpenAI.

This continues Avalon’s approach of supporting a range of leading AI models, allowing organisations to select the most appropriate model for each agent, workflow and use case.

What’s changed with GPT-6 Astra

OpenAI describes GPT-6 Astra as its most intelligent and aligned model to date, with significant improvements across computer use, browsing, software engineering, scientific reasoning and professional work.

Some of the most relevant improvements for enterprise AI include:

  • Stronger reasoning across complex, multi-step tasks
  • Significant improvements in computer and browser use
  • Improved software engineering and codebase understanding
  • Stronger performance across professional knowledge-work tasks
  • Improved ability to work across tools and applications
  • Better adherence to task boundaries and user intent

Astra is particularly interesting for agentic applications because the improvements extend beyond generating better responses. The model is increasingly capable of completing end-to-end workflows involving reasoning, tools, software and external systems.

Built for agent workflows

GPT-6 Astra supports many of the capabilities increasingly important to enterprise agents, including function calling, structured outputs, web and file search, computer use, MCP and Skills.

It also introduces capabilities such as asynchronous tool calling, allowing the model to continue useful work while an application is waiting for another tool to complete.

This direction aligns closely with Avalon’s agent architecture, where the underlying model works alongside tools, skills, enterprise knowledge, memory and governed execution.

How this fits within Zeaware Avalon

GPT-6 Astra is available as a selectable model within Zeaware Avalon and can be used across scenarios including:

  • Enterprise AI agents
  • Tool and skill-enabled workflows
  • Retrieval-based (RAG) solutions
  • Document and content generation
  • Research and analysis
  • Software and technical workflows
  • General-purpose AI interactions

Existing Avalon orchestration and governance patterns remain in place, allowing Astra to be introduced without redesigning the surrounding agent architecture.

This is important for enterprise deployments. The model is one component of an AI solution. Organisations still need appropriate grounding, tools, permissions, security controls, monitoring and governance around what an agent is allowed to do.

A significant step forward for enterprise agents

Perhaps the most interesting aspect of Astra is the continuing shift from models that primarily answer towards models that can reason and act.

For enterprise applications, this potentially means agents capable of completing increasingly substantial pieces of work rather than simply assisting users with individual steps.

The practical opportunity is not to give a model unrestricted autonomy, but to combine these stronger capabilities with well-defined tools, skills, knowledge and governance.

That is the model we are continuing to pursue with Zeaware Avalon.

What to consider

As with any major model release, organisations should evaluate Astra against their actual workloads rather than relying purely on benchmark results.

  • Testing existing agent workflows against Astra
  • Comparing quality with other models available within Avalon
  • Measuring end-to-end task performance rather than individual responses
  • Reviewing latency and token costs
  • Revalidating tool use and structured outputs
  • Applying appropriate controls to higher-autonomy workflows

Astra also represents a substantial increase in cybersecurity capability. OpenAI classifies it at the Critical capability level for cybersecurity under its Preparedness Framework, making appropriate safeguards particularly important for applications involving powerful tools or system access.

Availability

GPT-6 Astra support is now available in Zeaware Avalon.

Organisations using existing OpenAI, Anthropic or other supported models can evaluate Astra through Avalon’s normal model configuration and testing process without changing the broader agent architecture.

]]>> The AI Model Is Becoming the Least Important Part of Your AI Architecture https://live.cms.zeaware.net/zeaware/blog/the-ai-model-is-becoming-the-least-important-part-of-your-ai-architecture Tue, 08 Sep 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/the-ai-model-is-becoming-the-least-important-part-of-your-ai-architecture

For the last few years, much of the enterprise AI conversation has focused on models. Which model is smartest? Which has the largest context window? Which performs best on benchmarks?

Those questions still matter. But increasingly, they are not the questions that determine whether an enterprise AI system succeeds.

The model is becoming a replaceable component. The real architecture is everything around it.

Models Are Improving Faster Than Enterprise Architectures

The pace of model improvement is extraordinary. New models arrive with better reasoning, lower cost, faster inference and new capabilities. At the same time, organisations are gaining access to a growing range of commercial and open-source options.

This creates an important architectural implication: enterprises should avoid designing AI solutions around the assumption that one model, one provider or one model family will remain dominant.

A model that is the best choice today may not be the best choice in twelve months. In many environments, different models may also be appropriate for different workloads.

The long-term value therefore sits less in the model itself and more in the platform that determines how the model is used.

The Durable AI Architecture

A useful way to think about modern enterprise AI is as a set of layers:

Enterprise data → semantic context → agents → tools → governance → models

The model remains important, but it is only one part of a larger system.

The elements that are more likely to create durable enterprise value are the ones that connect the model to the organisation itself.

Context

AI systems need more than raw information. They need the right information for the task being performed.

That means understanding which customer, product, asset, case, policy, licence or business process the user is referring to, and then assembling the relevant knowledge around that entity.

Context may come from structured systems, documents, knowledge bases, APIs, data platforms and previous interactions. The challenge is not simply retrieving information. It is retrieving the right information in the right business context.

Tools

Useful enterprise AI does more than answer questions.

Agents increasingly need to query systems, retrieve records, update content, generate documents, initiate workflows or perform controlled actions. These capabilities are provided through tools and APIs rather than the model itself.

A capable model without access to enterprise tools is often little more than a sophisticated conversational interface.

Governance

Enterprise AI needs clear boundaries around what an agent can access, generate and execute.

Organisations need to know which data was used, which tools were invoked, which instructions were applied and what actions were taken. Permissions, auditability, policy enforcement and human approval become part of the architecture.

These controls are particularly important as AI systems move from answering questions to taking actions.

Orchestration

Not every task requires the same model, the same retrieval strategy or the same tools.

An orchestration layer can determine how a request should be handled: whether knowledge needs to be retrieved, whether a tool should be called, which model is most suitable and whether additional validation is required.

This is where much of the intelligence of an enterprise AI platform starts to move beyond the model itself.

Observability

When an AI system produces an unexpected result, organisations need to understand why.

What context was supplied? Which model was used? Which tools were called? What instructions were active? What source information influenced the response?

Without this visibility, AI systems can become difficult to operate, improve and govern at scale.

From RAG to an Enterprise Semantic Layer

Retrieval-augmented generation, or RAG, has been an important step in connecting language models to enterprise information. But basic RAG architectures often treat enterprise knowledge as a collection of documents to search.

That works for many use cases, but it is not always enough.

Consider a customer asking a question about a specific product. Relevant information may exist across a product information management system, installation manuals, warranty documents, knowledge articles, service cases, spreadsheets, technical drawings and internal systems.

Simply finding semantically similar text does not necessarily establish that all of those pieces of information relate to the same product, product range or customer context.

A stronger architecture introduces a semantic layer between enterprise data and the AI agent.

The semantic layer helps the system understand the entities and relationships that matter to the organisation: products, customers, assets, locations, contracts, policies, cases and other business concepts.

The agent can then reason from a grounded business entity outward, rather than searching through disconnected information and hoping that the right pieces are found.

Better AI Often Starts Outside the Model

When an enterprise AI application performs poorly, the instinct is often to try a more capable model.

Sometimes that helps. Often it does not.

The real problem may be incomplete data, weak retrieval, missing business relationships, poor tool design, insufficient instructions or a lack of governance.

A stronger model cannot reliably compensate for an architecture that provides the wrong context.

As models continue to improve, these surrounding architectural decisions become even more important because they determine how much value an organisation can extract from whichever model it chooses.

Design for Model Portability

Enterprises should increasingly treat models as a configurable part of the platform rather than a permanent dependency.

This does not mean that all models are equivalent. They are not. Different models have different strengths, cost profiles, hosting options and risk characteristics.

It means the enterprise architecture should make it possible to change models without redesigning the entire application.

A well-designed AI platform should allow organisations to adopt better models as they emerge, use different models for different workloads, and respond to changes in cost, regulation, sovereignty or vendor strategy.

The Competitive Advantage Moves Up the Stack

As foundation models become more capable and more widely available, access to a particular model is unlikely to remain a sustainable source of competitive advantage.

The differentiator will increasingly be how effectively an organisation connects those models to its own knowledge, systems and processes.

That means building strong enterprise context, reliable tools, clear governance, intelligent orchestration and comprehensive observability.

The winning enterprise AI architecture is therefore unlikely to be the one tied most closely to today's best model.

It will be the one that can continuously take advantage of better models without losing the enterprise knowledge, controls and capabilities built around them.

Models will continue to change. The architecture around them is where the durable value lives.

]]>

For the last few years, much of the enterprise AI conversation has focused on models. Which model is smartest? Which has the largest context window? Which performs best on benchmarks?

Those questions still matter. But increasingly, they are not the questions that determine whether an enterprise AI system succeeds.

The model is becoming a replaceable component. The real architecture is everything around it.

Models Are Improving Faster Than Enterprise Architectures

The pace of model improvement is extraordinary. New models arrive with better reasoning, lower cost, faster inference and new capabilities. At the same time, organisations are gaining access to a growing range of commercial and open-source options.

This creates an important architectural implication: enterprises should avoid designing AI solutions around the assumption that one model, one provider or one model family will remain dominant.

A model that is the best choice today may not be the best choice in twelve months. In many environments, different models may also be appropriate for different workloads.

The long-term value therefore sits less in the model itself and more in the platform that determines how the model is used.

The Durable AI Architecture

A useful way to think about modern enterprise AI is as a set of layers:

Enterprise data → semantic context → agents → tools → governance → models

The model remains important, but it is only one part of a larger system.

The elements that are more likely to create durable enterprise value are the ones that connect the model to the organisation itself.

Context

AI systems need more than raw information. They need the right information for the task being performed.

That means understanding which customer, product, asset, case, policy, licence or business process the user is referring to, and then assembling the relevant knowledge around that entity.

Context may come from structured systems, documents, knowledge bases, APIs, data platforms and previous interactions. The challenge is not simply retrieving information. It is retrieving the right information in the right business context.

Tools

Useful enterprise AI does more than answer questions.

Agents increasingly need to query systems, retrieve records, update content, generate documents, initiate workflows or perform controlled actions. These capabilities are provided through tools and APIs rather than the model itself.

A capable model without access to enterprise tools is often little more than a sophisticated conversational interface.

Governance

Enterprise AI needs clear boundaries around what an agent can access, generate and execute.

Organisations need to know which data was used, which tools were invoked, which instructions were applied and what actions were taken. Permissions, auditability, policy enforcement and human approval become part of the architecture.

These controls are particularly important as AI systems move from answering questions to taking actions.

Orchestration

Not every task requires the same model, the same retrieval strategy or the same tools.

An orchestration layer can determine how a request should be handled: whether knowledge needs to be retrieved, whether a tool should be called, which model is most suitable and whether additional validation is required.

This is where much of the intelligence of an enterprise AI platform starts to move beyond the model itself.

Observability

When an AI system produces an unexpected result, organisations need to understand why.

What context was supplied? Which model was used? Which tools were called? What instructions were active? What source information influenced the response?

Without this visibility, AI systems can become difficult to operate, improve and govern at scale.

From RAG to an Enterprise Semantic Layer

Retrieval-augmented generation, or RAG, has been an important step in connecting language models to enterprise information. But basic RAG architectures often treat enterprise knowledge as a collection of documents to search.

That works for many use cases, but it is not always enough.

Consider a customer asking a question about a specific product. Relevant information may exist across a product information management system, installation manuals, warranty documents, knowledge articles, service cases, spreadsheets, technical drawings and internal systems.

Simply finding semantically similar text does not necessarily establish that all of those pieces of information relate to the same product, product range or customer context.

A stronger architecture introduces a semantic layer between enterprise data and the AI agent.

The semantic layer helps the system understand the entities and relationships that matter to the organisation: products, customers, assets, locations, contracts, policies, cases and other business concepts.

The agent can then reason from a grounded business entity outward, rather than searching through disconnected information and hoping that the right pieces are found.

Better AI Often Starts Outside the Model

When an enterprise AI application performs poorly, the instinct is often to try a more capable model.

Sometimes that helps. Often it does not.

The real problem may be incomplete data, weak retrieval, missing business relationships, poor tool design, insufficient instructions or a lack of governance.

A stronger model cannot reliably compensate for an architecture that provides the wrong context.

As models continue to improve, these surrounding architectural decisions become even more important because they determine how much value an organisation can extract from whichever model it chooses.

Design for Model Portability

Enterprises should increasingly treat models as a configurable part of the platform rather than a permanent dependency.

This does not mean that all models are equivalent. They are not. Different models have different strengths, cost profiles, hosting options and risk characteristics.

It means the enterprise architecture should make it possible to change models without redesigning the entire application.

A well-designed AI platform should allow organisations to adopt better models as they emerge, use different models for different workloads, and respond to changes in cost, regulation, sovereignty or vendor strategy.

The Competitive Advantage Moves Up the Stack

As foundation models become more capable and more widely available, access to a particular model is unlikely to remain a sustainable source of competitive advantage.

The differentiator will increasingly be how effectively an organisation connects those models to its own knowledge, systems and processes.

That means building strong enterprise context, reliable tools, clear governance, intelligent orchestration and comprehensive observability.

The winning enterprise AI architecture is therefore unlikely to be the one tied most closely to today's best model.

It will be the one that can continuously take advantage of better models without losing the enterprise knowledge, controls and capabilities built around them.

Models will continue to change. The architecture around them is where the durable value lives.

]]>> Tony Bain Speaking at AI in Practice Panel https://live.cms.zeaware.net/zeaware/blog/deakin_ai_breakfast_panel_2026_event Tue, 21 Apr 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/deakin_ai_breakfast_panel_2026_event

Zeaware CEO Tony Bain will be a panelist at the upcoming event AI in Practice: Turning Hype into Measurable SME Impact.

The session brings together a range of speakers to discuss how small and medium-sized organisations are approaching AI, with a focus on practical use rather than theory.

About the session

The panel is centred on a common question - how to move from experimenting with AI to actually using it in day-to-day operations.

Topics are expected to include:

  • Where AI is currently being used in SMEs
  • Challenges in getting from pilot to production
  • What tends to work in practice
  • The role of data, workflows, and governance

Tony’s contribution

Tony will share perspectives based on Zeaware’s work building and deploying AI systems in business environments.

This includes a focus on:

  • Making AI work within existing processes
  • Keeping outputs reliable and consistent
  • Applying governance as part of the design

Event details

The event will be held as a breakfast panel session in Melbourne, bringing together practitioners and business leaders interested in practical AI adoption.

More information and registration is available here: View event details

]]>

Zeaware CEO Tony Bain will be a panelist at the upcoming event AI in Practice: Turning Hype into Measurable SME Impact.

The session brings together a range of speakers to discuss how small and medium-sized organisations are approaching AI, with a focus on practical use rather than theory.

About the session

The panel is centred on a common question - how to move from experimenting with AI to actually using it in day-to-day operations.

Topics are expected to include:

  • Where AI is currently being used in SMEs
  • Challenges in getting from pilot to production
  • What tends to work in practice
  • The role of data, workflows, and governance

Tony’s contribution

Tony will share perspectives based on Zeaware’s work building and deploying AI systems in business environments.

This includes a focus on:

  • Making AI work within existing processes
  • Keeping outputs reliable and consistent
  • Applying governance as part of the design

Event details

The event will be held as a breakfast panel session in Melbourne, bringing together practitioners and business leaders interested in practical AI adoption.

More information and registration is available here: View event details

]]>> Claude Opus 4.7 now supported in Zeaware Avalon https://live.cms.zeaware.net/zeaware/blog/anthropic_claude_opus_4_7_support Tue, 21 Apr 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/anthropic_claude_opus_4_7_support

Zeaware Avalon now supports Claude Opus 4.7, the latest model released by Anthropic.

This continues Zeaware Avalon’s approach of supporting a range of leading models, allowing organisations to select the most appropriate option for their specific use case.

What’s changed in Opus 4.7

According to Anthropic, Opus 4.7 includes updates across several areas:

  • Improvements to reasoning and multi-step task handling
  • Better performance in software engineering and agent-style workflows
  • More consistent instruction following
  • Enhancements to output quality and coherence
  • Ongoing updates to safety and model behaviour

As with all model updates, actual impact will depend on how the model is applied within a given solution.

How this fits within Zeaware Avalon

Opus 4.7 is available as a selectable model within Zeaware Avalon and can be used across:

  • Agent workflows
  • Retrieval-based (RAG) use cases
  • Document and content generation
  • General-purpose interactions

No changes are required to existing orchestration, governance, or integration patterns - model selection remains configurable at the agent level.

What to consider

  • Test against existing use cases to validate behaviour and output
  • Compare results with other available models where appropriate
  • Review cost, latency, and response characteristics

Model updates can introduce differences in behaviour, so validation in your specific context remains important.

Availability

Claude Opus 4.7 is now available in Zeaware Avalon environments.

If you are currently using Opus 4.6 or other models, you can evaluate 4.7 as part of your normal model selection and testing process.

]]>

Zeaware Avalon now supports Claude Opus 4.7, the latest model released by Anthropic.

This continues Zeaware Avalon’s approach of supporting a range of leading models, allowing organisations to select the most appropriate option for their specific use case.

What’s changed in Opus 4.7

According to Anthropic, Opus 4.7 includes updates across several areas:

  • Improvements to reasoning and multi-step task handling
  • Better performance in software engineering and agent-style workflows
  • More consistent instruction following
  • Enhancements to output quality and coherence
  • Ongoing updates to safety and model behaviour

As with all model updates, actual impact will depend on how the model is applied within a given solution.

How this fits within Zeaware Avalon

Opus 4.7 is available as a selectable model within Zeaware Avalon and can be used across:

  • Agent workflows
  • Retrieval-based (RAG) use cases
  • Document and content generation
  • General-purpose interactions

No changes are required to existing orchestration, governance, or integration patterns - model selection remains configurable at the agent level.

What to consider

  • Test against existing use cases to validate behaviour and output
  • Compare results with other available models where appropriate
  • Review cost, latency, and response characteristics

Model updates can introduce differences in behaviour, so validation in your specific context remains important.

Availability

Claude Opus 4.7 is now available in Zeaware Avalon environments.

If you are currently using Opus 4.6 or other models, you can evaluate 4.7 as part of your normal model selection and testing process.

]]>> Scaling Enterprise AI https://live.cms.zeaware.net/zeaware/blog/scaling_enterprise_ai Sun, 12 Apr 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/scaling_enterprise_ai

Over the past 12 to 18 months, most organisations have moved beyond asking whether they should use AI.

They have run pilots, built internal tools, tested use cases, and in many instances deployed early assistants and agents into real workflows. For many, those initiatives have delivered genuine value.

Yet for others, a consistent pattern has emerged. Progress often slows at the point where AI begins to matter most - when organisations attempt to move from isolated success into broader, operational use.

In many cases, this slowdown is not a failure of the technology. Rather, it reflects what can happen when a new capability like AI is introduced without a clear model for how it operates, scales, and integrates into the enterprise.

From capability to coordination

Early AI success is relatively contained. A model is applied to a specific task, connected to a dataset, and guided with a well-formed prompt. Within that boundary, outcomes can be impressive.

However, as soon as that capability is extended across teams, systems, and processes, the nature of the challenge changes.

It is no longer just about whether an agent can produce a useful response. It becomes a question of coordination:

  • Which systems can it access, and under what conditions
  • What data is appropriate and trusted for a given context
  • How decisions are validated before they are acted upon
  • How outputs are traced, explained, and reviewed
  • Where accountability ultimately sits

These are not new problems. They are the same questions enterprises have always had to answer when introducing new systems into core operations. AI simply brings them into sharper focus.

When AI becomes part of the operating model

As organisations push further, AI starts to take on a different role.

Agents are no longer confined to answering questions or generating content. They begin to interact with multiple systems, execute steps within workflows, and produce outputs that influence real business outcomes.

At that point, AI is no longer just a tool. It becomes part of the operating model.

And operating models require structure.

Without that structure, even highly capable AI introduces friction. Teams hesitate to trust outputs. Processes require additional layers of manual validation. Scaling becomes inconsistent, and governance is applied after the fact rather than built in from the start.

This is often where the gap between a successful pilot and a sustainable production deployment becomes most visible.

The limitation isn’t intelligence

A common response is to focus on improving the AI itself - refining prompts, enhancing retrieval, or adopting more advanced models.

Those steps can improve outcomes, but they rarely address the underlying constraint.

Because the challenge is not only about generating better answers. It is about controlling how those answers are formed, how they are validated, and how they are used within a broader process.

Without that level of control, capability alone does not translate into reliability.

The emergence of structure around AI

What is beginning to take shape across many organisations is a more deliberate approach to how AI operates within the enterprise.

For many, this has not yet been formally defined. However, together with our partners, we have been working to shape this more explicitly as organisations define how AI should operate, scale, and integrate within their environments.

This includes structure around:

  • How agents are defined, deployed, and allowed to interact
  • How tools and data sources are accessed and governed
  • How workflows are coordinated across systems
  • Where human touchpoints are introduced for validation and oversight
  • How policies and guardrails are applied consistently
  • How execution is recorded, reviewed, and understood

In many ways, this mirrors previous shifts in enterprise technology. Software development did not scale simply by writing better code. It scaled through the introduction of environments, pipelines, governance models, and operational discipline.

AI is now moving through a similar transition.

Where this becomes real

The inflection point is rarely technical. It is operational.

It is the moment when an organisation moves from proving that something can work, to relying on it to work consistently within a live process.

At that point, priorities change:

  • Consistency becomes more important than novelty
  • Traceability becomes as important as speed
  • Control becomes as important as capability

This is where many of the more meaningful decisions around AI are now being made.

A shift that is still unfolding

There remains a strong focus on models, benchmarks, and rapid capability improvements. That momentum will continue.

But increasingly, the more consequential work is happening elsewhere - in how organisations bring those capabilities into environments where reliability, accountability, and structure are required.

Not by constraining AI, but by giving it a framework within which it can operate effectively.

Across the organisations we work with, this shift is often being shaped in partnership with system integrators and advisory teams who are helping organisations define what that operating model should look like - how AI fits into processes, how governance is applied, and how accountability is maintained.

What is becoming equally important is how that operating model is then put into practice, consistently and at scale - particularly as the number of agents, tools, and interconnected processes begins to grow.

At Zeaware, this is where much of our focus sits. Ensuring that once those structures are defined, they can be operationalised in a way that is repeatable, observable, and aligned with how the organisation actually works. It is also a key driver behind how we continue to evolve Avalon.

More broadly, it reflects a shift that is becoming increasingly evident across the enterprise landscape.

One that is less about what AI can do, and more about how it is enabled to do it well.

]]>

Over the past 12 to 18 months, most organisations have moved beyond asking whether they should use AI.

They have run pilots, built internal tools, tested use cases, and in many instances deployed early assistants and agents into real workflows. For many, those initiatives have delivered genuine value.

Yet for others, a consistent pattern has emerged. Progress often slows at the point where AI begins to matter most - when organisations attempt to move from isolated success into broader, operational use.

In many cases, this slowdown is not a failure of the technology. Rather, it reflects what can happen when a new capability like AI is introduced without a clear model for how it operates, scales, and integrates into the enterprise.

From capability to coordination

Early AI success is relatively contained. A model is applied to a specific task, connected to a dataset, and guided with a well-formed prompt. Within that boundary, outcomes can be impressive.

However, as soon as that capability is extended across teams, systems, and processes, the nature of the challenge changes.

It is no longer just about whether an agent can produce a useful response. It becomes a question of coordination:

  • Which systems can it access, and under what conditions
  • What data is appropriate and trusted for a given context
  • How decisions are validated before they are acted upon
  • How outputs are traced, explained, and reviewed
  • Where accountability ultimately sits

These are not new problems. They are the same questions enterprises have always had to answer when introducing new systems into core operations. AI simply brings them into sharper focus.

When AI becomes part of the operating model

As organisations push further, AI starts to take on a different role.

Agents are no longer confined to answering questions or generating content. They begin to interact with multiple systems, execute steps within workflows, and produce outputs that influence real business outcomes.

At that point, AI is no longer just a tool. It becomes part of the operating model.

And operating models require structure.

Without that structure, even highly capable AI introduces friction. Teams hesitate to trust outputs. Processes require additional layers of manual validation. Scaling becomes inconsistent, and governance is applied after the fact rather than built in from the start.

This is often where the gap between a successful pilot and a sustainable production deployment becomes most visible.

The limitation isn’t intelligence

A common response is to focus on improving the AI itself - refining prompts, enhancing retrieval, or adopting more advanced models.

Those steps can improve outcomes, but they rarely address the underlying constraint.

Because the challenge is not only about generating better answers. It is about controlling how those answers are formed, how they are validated, and how they are used within a broader process.

Without that level of control, capability alone does not translate into reliability.

The emergence of structure around AI

What is beginning to take shape across many organisations is a more deliberate approach to how AI operates within the enterprise.

For many, this has not yet been formally defined. However, together with our partners, we have been working to shape this more explicitly as organisations define how AI should operate, scale, and integrate within their environments.

This includes structure around:

  • How agents are defined, deployed, and allowed to interact
  • How tools and data sources are accessed and governed
  • How workflows are coordinated across systems
  • Where human touchpoints are introduced for validation and oversight
  • How policies and guardrails are applied consistently
  • How execution is recorded, reviewed, and understood

In many ways, this mirrors previous shifts in enterprise technology. Software development did not scale simply by writing better code. It scaled through the introduction of environments, pipelines, governance models, and operational discipline.

AI is now moving through a similar transition.

Where this becomes real

The inflection point is rarely technical. It is operational.

It is the moment when an organisation moves from proving that something can work, to relying on it to work consistently within a live process.

At that point, priorities change:

  • Consistency becomes more important than novelty
  • Traceability becomes as important as speed
  • Control becomes as important as capability

This is where many of the more meaningful decisions around AI are now being made.

A shift that is still unfolding

There remains a strong focus on models, benchmarks, and rapid capability improvements. That momentum will continue.

But increasingly, the more consequential work is happening elsewhere - in how organisations bring those capabilities into environments where reliability, accountability, and structure are required.

Not by constraining AI, but by giving it a framework within which it can operate effectively.

Across the organisations we work with, this shift is often being shaped in partnership with system integrators and advisory teams who are helping organisations define what that operating model should look like - how AI fits into processes, how governance is applied, and how accountability is maintained.

What is becoming equally important is how that operating model is then put into practice, consistently and at scale - particularly as the number of agents, tools, and interconnected processes begins to grow.

At Zeaware, this is where much of our focus sits. Ensuring that once those structures are defined, they can be operationalised in a way that is repeatable, observable, and aligned with how the organisation actually works. It is also a key driver behind how we continue to evolve Avalon.

More broadly, it reflects a shift that is becoming increasingly evident across the enterprise landscape.

One that is less about what AI can do, and more about how it is enabled to do it well.

]]>> Fix the Data, Not the AI https://live.cms.zeaware.net/zeaware/blog/ai_data_quality Tue, 24 Mar 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/ai_data_quality Why the fastest path to better AI outcomes is improving your data - not your prompts

Most teams trying to improve AI performance follow the same path.

They tweak prompts.
They switch models.
They add more context.
They introduce validation layers.

It helps - but sometimes only marginally.

Because in many cases, the problem isn’t the AI.

It’s the data.

A common reason AI outputs fall short

Across enterprise deployments, we consistently see the same data issues:

  • Content that is duplicated or slightly inconsistent
  • Missing structure across documents
  • Conflicting information across sources
  • Important context buried in headings or formatting
  • Poorly chunked data that loses meaning when retrieved

When an AI system produces an answer that feels unreliable, it’s often doing exactly what it was asked to do - using the information available to it.

The issue is that the information itself is not reliable enough.

The anti-pattern: fixing outputs in the AI Assistant

A common response is to try and “correct” the AI at runtime.

This typically involves:

  • Prompt tuning loops
  • Adding more retrieval context
  • Rewriting answers before presenting them
  • Adding additional validation passes
  • Hardcoding "fix/workaround" prompts

These approaches treat the symptom, not the cause.

They can improve individual responses, but they don’t create a system that gets better over time.

In many cases, they also introduce:

  • Increased latency
  • Higher costs from additional model calls
  • More complexity in orchestration

And the underlying data problems remain unchanged.

POCs vs Production: where this approach breaks down

For early proof-of-concepts, working around data issues is often acceptable.

The goal at that stage is to demonstrate capability, not perfection. Prompt tuning, adding context, and refining responses can be enough to show what’s possible.

But this approach doesn’t scale.

As soon as you move toward pilot or production, the cracks become obvious:

  • Inconsistent answers across similar questions
  • Increasing reliance on complex prompts
  • Escalating cost and latency from additional model calls
  • Lack of confidence from users and stakeholders

At this point, continuing to work around data issues becomes a liability.

Production systems require a deliberate approach to data quality management.

A better approach: fix data at the source

The most effective improvements we’re seeing come from a different strategy entirely:

Improving the data itself.

Instead of trying to fix answers after they are generated, organisations are using AI to analyse and improve their knowledge sources offline.

This shifts the focus from:

“How do we get a better answer?”

to:

“How do we ensure the system is working from better information?”

The AI-driven data improvement loop

A more effective pattern is emerging.

AI is used not just for answering questions, but for continuously improving the content it relies on.

This typically involves:

1. Analysing existing content

AI reviews documents, product data, policies, and knowledge bases to identify:

  • duplication
  • inconsistencies
  • gaps in coverage
  • conflicting statements
  • poor structure or formatting

2. Proposing improvements

Rather than rewriting content blindly, the system proposes structured changes such as:

  • consolidating duplicate content
  • resolving inconsistencies
  • restructuring sections for clarity
  • enriching missing information

3. Presenting changes with context

Each proposed change includes:

  • before and after comparisons
  • supporting evidence
  • references to source material

Human-in-the-loop: control, not automation

Of course, enterprise data cannot be automatically rewritten without oversight.

This is where a human-in-the-loop process becomes critical.

Instead of fully automated changes, organisations implement a controlled workflow where:

  • proposed changes are reviewed
  • content can be edited before approval
  • decisions are explicitly approved or rejected
  • a full audit trail is maintained

This creates:

  • accountability
  • transparency
  • confidence in changes

And importantly, it ensures that improvements are deliberate, not accidental.

From static knowledge to continuously improving systems

When this approach is applied consistently, something important happens.

The system improves over time - not because the model changes, but because the data improves.

This leads to:

  • more consistent answers
  • reduced ambiguity
  • better retrieval outcomes
  • fewer edge cases
  • less reliance on prompt engineering

AI stops being something that needs constant correction, and starts becoming part of a broader knowledge improvement system.

The bottom line

If you want better AI outcomes, start by improving your data.

Not just once, but continuously.

Because the most effective AI systems aren’t the ones with the best prompts.

They’re the ones built on reliable, structured, and actively maintained knowledge.

]]> Why the fastest path to better AI outcomes is improving your data - not your prompts

Most teams trying to improve AI performance follow the same path.

They tweak prompts.
They switch models.
They add more context.
They introduce validation layers.

It helps - but sometimes only marginally.

Because in many cases, the problem isn’t the AI.

It’s the data.

A common reason AI outputs fall short

Across enterprise deployments, we consistently see the same data issues:

  • Content that is duplicated or slightly inconsistent
  • Missing structure across documents
  • Conflicting information across sources
  • Important context buried in headings or formatting
  • Poorly chunked data that loses meaning when retrieved

When an AI system produces an answer that feels unreliable, it’s often doing exactly what it was asked to do - using the information available to it.

The issue is that the information itself is not reliable enough.

The anti-pattern: fixing outputs in the AI Assistant

A common response is to try and “correct” the AI at runtime.

This typically involves:

  • Prompt tuning loops
  • Adding more retrieval context
  • Rewriting answers before presenting them
  • Adding additional validation passes
  • Hardcoding "fix/workaround" prompts

These approaches treat the symptom, not the cause.

They can improve individual responses, but they don’t create a system that gets better over time.

In many cases, they also introduce:

  • Increased latency
  • Higher costs from additional model calls
  • More complexity in orchestration

And the underlying data problems remain unchanged.

POCs vs Production: where this approach breaks down

For early proof-of-concepts, working around data issues is often acceptable.

The goal at that stage is to demonstrate capability, not perfection. Prompt tuning, adding context, and refining responses can be enough to show what’s possible.

But this approach doesn’t scale.

As soon as you move toward pilot or production, the cracks become obvious:

  • Inconsistent answers across similar questions
  • Increasing reliance on complex prompts
  • Escalating cost and latency from additional model calls
  • Lack of confidence from users and stakeholders

At this point, continuing to work around data issues becomes a liability.

Production systems require a deliberate approach to data quality management.

A better approach: fix data at the source

The most effective improvements we’re seeing come from a different strategy entirely:

Improving the data itself.

Instead of trying to fix answers after they are generated, organisations are using AI to analyse and improve their knowledge sources offline.

This shifts the focus from:

“How do we get a better answer?”

to:

“How do we ensure the system is working from better information?”

The AI-driven data improvement loop

A more effective pattern is emerging.

AI is used not just for answering questions, but for continuously improving the content it relies on.

This typically involves:

1. Analysing existing content

AI reviews documents, product data, policies, and knowledge bases to identify:

  • duplication
  • inconsistencies
  • gaps in coverage
  • conflicting statements
  • poor structure or formatting

2. Proposing improvements

Rather than rewriting content blindly, the system proposes structured changes such as:

  • consolidating duplicate content
  • resolving inconsistencies
  • restructuring sections for clarity
  • enriching missing information

3. Presenting changes with context

Each proposed change includes:

  • before and after comparisons
  • supporting evidence
  • references to source material

Human-in-the-loop: control, not automation

Of course, enterprise data cannot be automatically rewritten without oversight.

This is where a human-in-the-loop process becomes critical.

Instead of fully automated changes, organisations implement a controlled workflow where:

  • proposed changes are reviewed
  • content can be edited before approval
  • decisions are explicitly approved or rejected
  • a full audit trail is maintained

This creates:

  • accountability
  • transparency
  • confidence in changes

And importantly, it ensures that improvements are deliberate, not accidental.

From static knowledge to continuously improving systems

When this approach is applied consistently, something important happens.

The system improves over time - not because the model changes, but because the data improves.

This leads to:

  • more consistent answers
  • reduced ambiguity
  • better retrieval outcomes
  • fewer edge cases
  • less reliance on prompt engineering

AI stops being something that needs constant correction, and starts becoming part of a broader knowledge improvement system.

The bottom line

If you want better AI outcomes, start by improving your data.

Not just once, but continuously.

Because the most effective AI systems aren’t the ones with the best prompts.

They’re the ones built on reliable, structured, and actively maintained knowledge.

]]>> Prompt Repeat - a simple option with measurable impact https://live.cms.zeaware.net/zeaware/blog/avalon-prompt-repeat Fri, 20 Feb 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/avalon-prompt-repeat

We’ve just added a Prompt Repeat option to Zeaware Avalon, allowing instructions to be selectively repeated during prompt assembly. In the right situations, this can help improve accuracy, consistency, and predictability, without changing models, pipelines, or architectures.

Where Prompt Repeat helps most

  • Classification
  • Structured / fixed-schema extraction
  • Intent detection
  • Rule-constrained responses
  • Validation and completeness checks

This feature is informed by recent research, including “Prompt Repetition Improves Non-Reasoning LLMs” (arXiv:2512.14982), which reported 3–10% absolute accuracy improvements on non-reasoning tasks simply by repeating the user query (e.g. Q → Q Q). Gains were strongest where instruction adherence matters more than step-by-step reasoning, and were achieved without modifying the underlying model.

Designed for controlled use

In Zeaware Avalon, Prompt Repeat is configurable (applied only where it helps), scoped (instructions, not entire contexts), and governed (designed to preserve deterministic outputs).

It’s a useful reminder that improving AI systems isn’t always about bigger models or more complexity, sometimes it’s about execution of simple, evidence-backed ideas.

Note: Prompt Repeat increases input token usage and may result in higher token consumption. As with all prompt-level techniques, observed improvements can vary by model, task, and data characteristics. Prompt Repeat is therefore configurable and should be applied selectively where it provides measurable benefit.

]]>

We’ve just added a Prompt Repeat option to Zeaware Avalon, allowing instructions to be selectively repeated during prompt assembly. In the right situations, this can help improve accuracy, consistency, and predictability, without changing models, pipelines, or architectures.

Where Prompt Repeat helps most

  • Classification
  • Structured / fixed-schema extraction
  • Intent detection
  • Rule-constrained responses
  • Validation and completeness checks

This feature is informed by recent research, including “Prompt Repetition Improves Non-Reasoning LLMs” (arXiv:2512.14982), which reported 3–10% absolute accuracy improvements on non-reasoning tasks simply by repeating the user query (e.g. Q → Q Q). Gains were strongest where instruction adherence matters more than step-by-step reasoning, and were achieved without modifying the underlying model.

Designed for controlled use

In Zeaware Avalon, Prompt Repeat is configurable (applied only where it helps), scoped (instructions, not entire contexts), and governed (designed to preserve deterministic outputs).

It’s a useful reminder that improving AI systems isn’t always about bigger models or more complexity, sometimes it’s about execution of simple, evidence-backed ideas.

Note: Prompt Repeat increases input token usage and may result in higher token consumption. As with all prompt-level techniques, observed improvements can vary by model, task, and data characteristics. Prompt Repeat is therefore configurable and should be applied selectively where it provides measurable benefit.

]]>> From Interoperability to Agency: A Zeaware View on Practical AI Sovereignty https://live.cms.zeaware.net/zeaware/blog/an_avalon_view_on_practical_ai_sovereignty Fri, 20 Feb 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/an_avalon_view_on_practical_ai_sovereignty

This article responds to and builds on Why AI Sovereignty Depends on Interoperability Standards by Eileen Donahoe and Konstantinos Komaitis, published by Tech Policy Press on 17 February 2026.

AI sovereignty is no longer a fringe concern reserved for national security agencies or technology ministries. As artificial intelligence becomes embedded in public services, regulatory decision-making, and critical infrastructure, sovereignty has become a practical question: who ultimately controls how these systems behave, evolve, and can be changed.

Donahoe and Komaitis argue persuasively that sovereignty in the AI era is exercised less through ownership of models and more through interoperable standards, particularly at the interfaces, orchestration layers, and governance mechanisms that shape how systems operate.

Where this discussion can be extended is in how interoperability becomes operational rather than aspirational.

Interoperability is not about language

A subtle assumption often appears in AI discussions: because large language models operate in shared natural languages, particularly English, interoperability is largely implicit.

In practice, this is not the case.

Models differ in ways that matter deeply for governance and control, including tool invocation semantics, refusal behaviour, confidence expression, logging formats, memory handling, and agent execution patterns.

Without a platform layer that absorbs these differences, they become embedded dependencies. A system may technically be able to switch models, but doing so can alter behaviour, auditability, and risk posture in ways that are difficult to predict or control.

English is a shared interface for users, not a control surface for systems.

Sovereignty is the ability to change models without changing the system

True AI sovereignty does not require exclusive ownership of models. It requires the ability to replace them without losing control.

This is the design principle behind Zeaware Avalon.

Zeaware Avalon abstracts model-specific behaviour behind a consistent orchestration and governance layer so that, from a user and operator perspective, prompts remain stable, agent logic remains stable, policy constraints remain stable, and audit and decision records remain comparable.

The model becomes a replaceable execution component, not the defining feature of the system.

This matters for governments and regulated organisations because procurement strategies change, risk tolerances evolve, regulatory obligations shift, and geopolitical or supply-chain constraints emerge.

If changing a model requires rewriting prompts, re-authoring workflows, or rebuilding governance processes, sovereignty has already been compromised.

Governance must survive model changes

One of the risks highlighted implicitly in the interoperability debate is that governance is often embedded inside models rather than enforced above them.

This includes safety rules embedded in weights, opaque refusal logic, undocumented prioritisation of values, and uninspectable moderation thresholds.

Zeaware Avalon deliberately moves governance up the stack into explicit, enforceable system controls. These include policy-aligned constraints, deterministic validation rules, structured decision checkpoints, human-in-the-loop escalation, and replayable execution histories.

Because these controls sit outside the model, they persist regardless of which model is used.

Interoperability without governance persistence is only surface-level sovereignty.

Agent orchestration is the real control plane

As AI systems become increasingly agentic, invoking tools, accessing data, and acting autonomously, the strategic question shifts.

Who controls the orchestration logic.

From an Zeaware Avalon perspective, orchestration is treated as a first-class, inspectable, governed layer.

Agents follow defined execution paths. Tool access is policy-gated. Decisions can pause, escalate, or be overridden. Execution histories can be audited, reviewed, and replayed.

This ensures behavioural authority remains with the system owner, whether a government department or a regulated enterprise, rather than being delegated to a model provider by default.

Interoperability as experienced sovereignty

A useful test of sovereignty is not theoretical compliance, but lived operational experience.

In practice, interoperable sovereignty means models can be trialled and compared safely, regulatory changes do not require replatforming, better models can be adopted without re-authoring systems, and vendors can be exited without losing institutional knowledge.

This is what sovereignty looks like when it is experienced, not just asserted.

A complementary conclusion

The Tech Policy Press article is right. Open standards, modularity, and coordination matter. But standards alone do not deliver sovereignty.

Sovereignty emerges when model differences are abstracted, governance is externalised from models, orchestration is controlled and inspectable, and exit is operationally real rather than merely contractual.

Platforms that make interoperability practical rather than theoretical are where AI sovereignty is actually realised.

Sovereignty is no longer something governments must wait for global consensus to provide. It is something they can design into systems today, deliberately, visibly, and on their own terms.

]]>

This article responds to and builds on Why AI Sovereignty Depends on Interoperability Standards by Eileen Donahoe and Konstantinos Komaitis, published by Tech Policy Press on 17 February 2026.

AI sovereignty is no longer a fringe concern reserved for national security agencies or technology ministries. As artificial intelligence becomes embedded in public services, regulatory decision-making, and critical infrastructure, sovereignty has become a practical question: who ultimately controls how these systems behave, evolve, and can be changed.

Donahoe and Komaitis argue persuasively that sovereignty in the AI era is exercised less through ownership of models and more through interoperable standards, particularly at the interfaces, orchestration layers, and governance mechanisms that shape how systems operate.

Where this discussion can be extended is in how interoperability becomes operational rather than aspirational.

Interoperability is not about language

A subtle assumption often appears in AI discussions: because large language models operate in shared natural languages, particularly English, interoperability is largely implicit.

In practice, this is not the case.

Models differ in ways that matter deeply for governance and control, including tool invocation semantics, refusal behaviour, confidence expression, logging formats, memory handling, and agent execution patterns.

Without a platform layer that absorbs these differences, they become embedded dependencies. A system may technically be able to switch models, but doing so can alter behaviour, auditability, and risk posture in ways that are difficult to predict or control.

English is a shared interface for users, not a control surface for systems.

Sovereignty is the ability to change models without changing the system

True AI sovereignty does not require exclusive ownership of models. It requires the ability to replace them without losing control.

This is the design principle behind Zeaware Avalon.

Zeaware Avalon abstracts model-specific behaviour behind a consistent orchestration and governance layer so that, from a user and operator perspective, prompts remain stable, agent logic remains stable, policy constraints remain stable, and audit and decision records remain comparable.

The model becomes a replaceable execution component, not the defining feature of the system.

This matters for governments and regulated organisations because procurement strategies change, risk tolerances evolve, regulatory obligations shift, and geopolitical or supply-chain constraints emerge.

If changing a model requires rewriting prompts, re-authoring workflows, or rebuilding governance processes, sovereignty has already been compromised.

Governance must survive model changes

One of the risks highlighted implicitly in the interoperability debate is that governance is often embedded inside models rather than enforced above them.

This includes safety rules embedded in weights, opaque refusal logic, undocumented prioritisation of values, and uninspectable moderation thresholds.

Zeaware Avalon deliberately moves governance up the stack into explicit, enforceable system controls. These include policy-aligned constraints, deterministic validation rules, structured decision checkpoints, human-in-the-loop escalation, and replayable execution histories.

Because these controls sit outside the model, they persist regardless of which model is used.

Interoperability without governance persistence is only surface-level sovereignty.

Agent orchestration is the real control plane

As AI systems become increasingly agentic, invoking tools, accessing data, and acting autonomously, the strategic question shifts.

Who controls the orchestration logic.

From an Zeaware Avalon perspective, orchestration is treated as a first-class, inspectable, governed layer.

Agents follow defined execution paths. Tool access is policy-gated. Decisions can pause, escalate, or be overridden. Execution histories can be audited, reviewed, and replayed.

This ensures behavioural authority remains with the system owner, whether a government department or a regulated enterprise, rather than being delegated to a model provider by default.

Interoperability as experienced sovereignty

A useful test of sovereignty is not theoretical compliance, but lived operational experience.

In practice, interoperable sovereignty means models can be trialled and compared safely, regulatory changes do not require replatforming, better models can be adopted without re-authoring systems, and vendors can be exited without losing institutional knowledge.

This is what sovereignty looks like when it is experienced, not just asserted.

A complementary conclusion

The Tech Policy Press article is right. Open standards, modularity, and coordination matter. But standards alone do not deliver sovereignty.

Sovereignty emerges when model differences are abstracted, governance is externalised from models, orchestration is controlled and inspectable, and exit is operationally real rather than merely contractual.

Platforms that make interoperability practical rather than theoretical are where AI sovereignty is actually realised.

Sovereignty is no longer something governments must wait for global consensus to provide. It is something they can design into systems today, deliberately, visibly, and on their own terms.

]]>> Anthropic Claude Sonnet 4.6 Support in Zeaware Avalon https://live.cms.zeaware.net/zeaware/blog/anthropic_claude_sonnet_4_6_support Fri, 20 Feb 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/anthropic_claude_sonnet_4_6_support

We’re pleased to announce native for Anthropic Claude Sonnet 4.6 in Zeaware Avalon. Sonnet 4.6 is Anthropic’s best combination of speed and intelligence - designed for production use cases that need strong reasoning and coding performance without the cost profile of a flagship model.

With this release, Claude Sonnet 4.6 is now available as a first-class Model Service within Zeaware Avalon.

Why Sonnet 4.6

Sonnet 4.6 is a practical default for many enterprise workloads - particularly where you want high-quality reasoning and coding, fast iteration, and predictable operating cost. Anthropic positions Sonnet 4.6 as the “best combination of speed and intelligence”.

It also introduces modern capability controls like adaptive thinking and the effort parameter, enabling you to tune performance vs cost depending on the scenario.

How Sonnet 4.6 Fits into Zeaware Avalon

In Zeaware Avalon, models are registered and managed as services.

This means Claude Sonnet 4.6 is added the same way you would add any other AI capability in the platform:

  • Register Anthropic Claude Sonnet 4.6 as a Model Service
  • Assign it to agents, workflows, or environments as required
  • Control where and how it is used via Zeaware Avalon’s governance and configuration layers

Where Sonnet 4.6 Adds Value

Sonnet 4.6 is particularly well suited to high-volume or always-on workloads where cost efficiency matters, but you still need strong capability:

  • Agent workflows and orchestration - multi-step tasks where consistent instruction-following and reliable reasoning matter. 
  • Coding and technical assistants - generation, refactoring, debugging, and analysis of larger codebases.
  • Long-context enterprise tasks - analysis across long documents, policies, procedures, or large knowledge packs (within supported context limits).
  • Structured extraction and classification - repeatable outputs where you want an effective balance of speed, quality, and operating cost.

Available Now in Zeaware Avalon

Claude Sonnet 4.6 can now be configured as a Model Service in Zeaware Avalon and applied selectively across agents and workflows based on capability, governance needs, and cost profile.

]]>

We’re pleased to announce native for Anthropic Claude Sonnet 4.6 in Zeaware Avalon. Sonnet 4.6 is Anthropic’s best combination of speed and intelligence - designed for production use cases that need strong reasoning and coding performance without the cost profile of a flagship model.

With this release, Claude Sonnet 4.6 is now available as a first-class Model Service within Zeaware Avalon.

Why Sonnet 4.6

Sonnet 4.6 is a practical default for many enterprise workloads - particularly where you want high-quality reasoning and coding, fast iteration, and predictable operating cost. Anthropic positions Sonnet 4.6 as the “best combination of speed and intelligence”.

It also introduces modern capability controls like adaptive thinking and the effort parameter, enabling you to tune performance vs cost depending on the scenario.

How Sonnet 4.6 Fits into Zeaware Avalon

In Zeaware Avalon, models are registered and managed as services.

This means Claude Sonnet 4.6 is added the same way you would add any other AI capability in the platform:

  • Register Anthropic Claude Sonnet 4.6 as a Model Service
  • Assign it to agents, workflows, or environments as required
  • Control where and how it is used via Zeaware Avalon’s governance and configuration layers

Where Sonnet 4.6 Adds Value

Sonnet 4.6 is particularly well suited to high-volume or always-on workloads where cost efficiency matters, but you still need strong capability:

  • Agent workflows and orchestration - multi-step tasks where consistent instruction-following and reliable reasoning matter. 
  • Coding and technical assistants - generation, refactoring, debugging, and analysis of larger codebases.
  • Long-context enterprise tasks - analysis across long documents, policies, procedures, or large knowledge packs (within supported context limits).
  • Structured extraction and classification - repeatable outputs where you want an effective balance of speed, quality, and operating cost.

Available Now in Zeaware Avalon

Claude Sonnet 4.6 can now be configured as a Model Service in Zeaware Avalon and applied selectively across agents and workflows based on capability, governance needs, and cost profile.

]]>> Anthropic Claude Opus 4.6 Support in Zeaware Avalon https://live.cms.zeaware.net/zeaware/blog/anthropic_claude_opus_4_6_support Tue, 10 Feb 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/anthropic_claude_opus_4_6_support

We’re pleased to announce native support for Anthropic Claude Opus 4.6 in Zeaware Avalon.

Opus 4.6 is Anthropic’s most capable general-purpose model to date, with strong performance across complex reasoning, long-context tasks, agentic workflows, and large-scale codebases. With this release, Opus 4.6 is now available as a first-class Model Service within Zeaware Avalon.

What is Claude Opus 4.6?

Claude Opus 4.6 is Anthropic’s flagship model, designed for demanding enterprise and knowledge-intensive workloads. It builds on earlier Opus versions with improvements in:

  • Multi-step reasoning and planning
  • Long-context comprehension and recall
  • Complex document analysis and synthesis
  • Agent-driven workflows and tool use
  • High-quality code generation and refactoring

These capabilities make Opus 4.6 well suited to regulated environments, research workflows, and applications where accuracy, consistency, and sustained reasoning matter.

How Opus 4.6 Fits into Avalon

In Zeaware Avalon, models are registered and managed as services.

This means Claude Opus 4.6 is added the same way you would add any other AI capability in the platform:

  • Register Anthropic Claude Opus 4.6 as a Model Service
  • Assign it to agents, workflows, or environments as required
  • Control where and how it is used via Zeaware Avalon’s governance and configuration layers

Once registered, Opus 4.6 becomes available alongside other supported models, allowing teams to select the right model for each task without changing application logic.

Why This Matters

Treating models as services enables a few important things:

  • Model choice becomes a configuration decision, not a code change
  • Different models can be used for different intents, workflows, or risk profiles
  • New models can be introduced safely, without disrupting existing applications
  • Governance and auditability are preserved as models evolve

For teams building agentic systems, document analysis tools, or high-context assistants, Opus 4.6 offers another strong option - particularly where long-running reasoning and document fidelity are critical.

We’ll continue to expand Zeaware Avalon’s model support so customers can take advantage of the latest capabilities from leading AI providers - while keeping control, governance, and flexibility at the centre of the platform.

]]>

We’re pleased to announce native support for Anthropic Claude Opus 4.6 in Zeaware Avalon.

Opus 4.6 is Anthropic’s most capable general-purpose model to date, with strong performance across complex reasoning, long-context tasks, agentic workflows, and large-scale codebases. With this release, Opus 4.6 is now available as a first-class Model Service within Zeaware Avalon.

What is Claude Opus 4.6?

Claude Opus 4.6 is Anthropic’s flagship model, designed for demanding enterprise and knowledge-intensive workloads. It builds on earlier Opus versions with improvements in:

  • Multi-step reasoning and planning
  • Long-context comprehension and recall
  • Complex document analysis and synthesis
  • Agent-driven workflows and tool use
  • High-quality code generation and refactoring

These capabilities make Opus 4.6 well suited to regulated environments, research workflows, and applications where accuracy, consistency, and sustained reasoning matter.

How Opus 4.6 Fits into Avalon

In Zeaware Avalon, models are registered and managed as services.

This means Claude Opus 4.6 is added the same way you would add any other AI capability in the platform:

  • Register Anthropic Claude Opus 4.6 as a Model Service
  • Assign it to agents, workflows, or environments as required
  • Control where and how it is used via Zeaware Avalon’s governance and configuration layers

Once registered, Opus 4.6 becomes available alongside other supported models, allowing teams to select the right model for each task without changing application logic.

Why This Matters

Treating models as services enables a few important things:

  • Model choice becomes a configuration decision, not a code change
  • Different models can be used for different intents, workflows, or risk profiles
  • New models can be introduced safely, without disrupting existing applications
  • Governance and auditability are preserved as models evolve

For teams building agentic systems, document analysis tools, or high-context assistants, Opus 4.6 offers another strong option - particularly where long-running reasoning and document fidelity are critical.

We’ll continue to expand Zeaware Avalon’s model support so customers can take advantage of the latest capabilities from leading AI providers - while keeping control, governance, and flexibility at the centre of the platform.

]]>> Create a Legal Document Search Engine with Zeaware Avalon and Isaacus Kanon 2 Embedder https://live.cms.zeaware.net/zeaware/blog/legal_document_search_engine_kanon_2_embedder Mon, 09 Feb 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/legal_document_search_engine_kanon_2_embedder

We build real-world AI tooling for business users - tools that help legal teams, compliance professionals, and knowledge workers make sense of complex content at scale. Zeaware Avalon ships with inbuilt services for ingestion, chunking, indexing, and assistant patterns, while still making it easy to integrate best-in-class external services via direct connectors, MCP, and standard protocols.

What’s New: Native Kanon 2 Embedder Support in Avalon

Zeaware Avalon now includes native support for Isaacus Kanon 2 Embedder. That means Kanon 2 can be registered as a first-class embedding service and used directly in Avalon’s Knowledge indexing flows - including vector and hybrid retrieval.

This is not a one-off integration. Kanon 2 is now part of our regular test regime, so it is supported and validated alongside the rest of Avalon’s indexing and retrieval stack.

Why Kanon 2 Embedder Matters for Legal Search

Isaacus’ Kanon 2 Embedder is a purpose-built legal embedding model designed to improve relevance for legal retrieval tasks like clause search, case law similarity, regulatory interpretation, and document discovery. For more background, see the Isaacus site: https://isaacus.com/.

We also love that Isaacus is an Australian AI company. Australia’s AI innovation community is strong and practical, and as another Australian AI company, Zeaware is proud to support local capability that competes globally.

Step-by-Step: Use Kanon 2 with Azure AI Search in Avalon

In this walkthrough, we’ll build a legal document search engine using Azure AI Search for the index and Kanon 2 Embedder for vector embeddings. The result is a Knowledge Index that supports semantic vector ranking, keyword matching (BM25), and deterministic filtering.

  1. Register Azure AI Search and Kanon 2 in Avalon Services

    Create or select your Azure AI Search service in the Azure Portal, and obtain the required API key(s). Separately, obtain your Kanon 2 Embedder API key from the Isaacus portal.

    In Avalon, navigate to Services and register:

    • Azure AI Search as a Search Service
    • Isaacus Kanon 2 Embedder as an Embedding Service
    @{#imageembed("image_add_isaccus")}

  2. Create a Vector (Hybrid) Knowledge Index

    In Avalon, go to Knowledge and select Add an Index. Choose Azure AI Search as the search service.

    • Select Vector Index (or Hybrid, if you want vector + keyword together)
    • Select Kanon 2 as the embedding service

    Avalon will configure the index so document chunks are embedded using Kanon 2 and stored in Azure AI Search for fast similarity search.

    @{#imageembed("image_add_index_kanon_2")}

  3. Add Documents (Quick Method)

    To index documents quickly, browse to your storage location in Avalon’s file browser, then right-click files or folders and choose Split and Index.

    Avalon will split content into chunks, generate Kanon 2 embeddings, and push the content + vectors into Azure AI Search.

    Production note: for a real deployment you would typically create a workflow with a file detection trigger and automate the split + index pipeline, rather than indexing manually.

    @{#imageembed("image_add_document_kanon2")}

  4. Use the Knowledge Index in Assistants and Search Experiences

    Once indexed, your Knowledge Index can be used across different assistant patterns in Avalon, including:

    • Chat Apps for fast document Q&A (RAG-style retrieval)
    • Search Assistants for retrieving documents based on criteria
    • Task-focused AI assistants that perform structured discovery jobs

    Example task: find all contracts with uncapped liability - combining semantic similarity (Kanon 2 vectors), keyword matching (BM25), and deterministic filters (document type, jurisdiction, date ranges, parties, and more).

    @{#imageembed("image_search_document_kanon2")}

Summary

Zeaware Avalon’s native support for Isaacus Kanon 2 Embedder gives teams a practical, enterprise-ready way to build high-quality legal search and assistant experiences - without needing to stitch together fragile glue code.

  • Native integration - supported and validated as part of Avalon’s regular test regime
  • Stronger legal retrieval - purpose-built embeddings tuned for legal language and intent
  • Hybrid retrieval - combine Kanon 2 vectors with Azure AI Search keyword search and filtering
  • Australian innovation - proud to support another Australian AI company building globally competitive capability

If you want to see this pattern applied to your own corpus (contracts, policies, case law, standards, or compliance documents), Zeaware Avalon makes it straightforward to go from documents to a working, governed search experience.

]]>

We build real-world AI tooling for business users - tools that help legal teams, compliance professionals, and knowledge workers make sense of complex content at scale. Zeaware Avalon ships with inbuilt services for ingestion, chunking, indexing, and assistant patterns, while still making it easy to integrate best-in-class external services via direct connectors, MCP, and standard protocols.

What’s New: Native Kanon 2 Embedder Support in Avalon

Zeaware Avalon now includes native support for Isaacus Kanon 2 Embedder. That means Kanon 2 can be registered as a first-class embedding service and used directly in Avalon’s Knowledge indexing flows - including vector and hybrid retrieval.

This is not a one-off integration. Kanon 2 is now part of our regular test regime, so it is supported and validated alongside the rest of Avalon’s indexing and retrieval stack.

Why Kanon 2 Embedder Matters for Legal Search

Isaacus’ Kanon 2 Embedder is a purpose-built legal embedding model designed to improve relevance for legal retrieval tasks like clause search, case law similarity, regulatory interpretation, and document discovery. For more background, see the Isaacus site: https://isaacus.com/.

We also love that Isaacus is an Australian AI company. Australia’s AI innovation community is strong and practical, and as another Australian AI company, Zeaware is proud to support local capability that competes globally.

Step-by-Step: Use Kanon 2 with Azure AI Search in Avalon

In this walkthrough, we’ll build a legal document search engine using Azure AI Search for the index and Kanon 2 Embedder for vector embeddings. The result is a Knowledge Index that supports semantic vector ranking, keyword matching (BM25), and deterministic filtering.

  1. Register Azure AI Search and Kanon 2 in Avalon Services

    Create or select your Azure AI Search service in the Azure Portal, and obtain the required API key(s). Separately, obtain your Kanon 2 Embedder API key from the Isaacus portal.

    In Avalon, navigate to Services and register:

    • Azure AI Search as a Search Service
    • Isaacus Kanon 2 Embedder as an Embedding Service
    @{#imageembed("image_add_isaccus")}

  2. Create a Vector (Hybrid) Knowledge Index

    In Avalon, go to Knowledge and select Add an Index. Choose Azure AI Search as the search service.

    • Select Vector Index (or Hybrid, if you want vector + keyword together)
    • Select Kanon 2 as the embedding service

    Avalon will configure the index so document chunks are embedded using Kanon 2 and stored in Azure AI Search for fast similarity search.

    @{#imageembed("image_add_index_kanon_2")}

  3. Add Documents (Quick Method)

    To index documents quickly, browse to your storage location in Avalon’s file browser, then right-click files or folders and choose Split and Index.

    Avalon will split content into chunks, generate Kanon 2 embeddings, and push the content + vectors into Azure AI Search.

    Production note: for a real deployment you would typically create a workflow with a file detection trigger and automate the split + index pipeline, rather than indexing manually.

    @{#imageembed("image_add_document_kanon2")}

  4. Use the Knowledge Index in Assistants and Search Experiences

    Once indexed, your Knowledge Index can be used across different assistant patterns in Avalon, including:

    • Chat Apps for fast document Q&A (RAG-style retrieval)
    • Search Assistants for retrieving documents based on criteria
    • Task-focused AI assistants that perform structured discovery jobs

    Example task: find all contracts with uncapped liability - combining semantic similarity (Kanon 2 vectors), keyword matching (BM25), and deterministic filters (document type, jurisdiction, date ranges, parties, and more).

    @{#imageembed("image_search_document_kanon2")}

Summary

Zeaware Avalon’s native support for Isaacus Kanon 2 Embedder gives teams a practical, enterprise-ready way to build high-quality legal search and assistant experiences - without needing to stitch together fragile glue code.

  • Native integration - supported and validated as part of Avalon’s regular test regime
  • Stronger legal retrieval - purpose-built embeddings tuned for legal language and intent
  • Hybrid retrieval - combine Kanon 2 vectors with Azure AI Search keyword search and filtering
  • Australian innovation - proud to support another Australian AI company building globally competitive capability

If you want to see this pattern applied to your own corpus (contracts, policies, case law, standards, or compliance documents), Zeaware Avalon makes it straightforward to go from documents to a working, governed search experience.

]]>> Tool Exfiltration Attacks, GenAI, and Why Control Matters https://live.cms.zeaware.net/zeaware/blog/tool_exfiltration_and_why_control_matters Mon, 05 Jan 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/tool_exfiltration_and_why_control_matters Recent discussion around tool exfiltration and indirect prompt injection attacks in Generative AI systems has raised valid concerns - particularly where platforms unexpectedly invoke tools or actions as a result of untrusted input. These concerns are worth taking seriously. But they are often framed in a way that blurs the distinction between the model, the application, and the execution platform.

This post clarifies that distinction, explains where risk actually lives, and outlines how enterprise AI systems can be designed to limit exposure through explicit control.

The LLM Is Just a Model

At its core, a large language model (LLM) does one thing: it takes an input and generates an output. That output may be plain text for a human, or it may be a structured response that an application interprets as a suggestion to call a tool.

The model itself has no execution capability, no awareness of trust boundaries, and no understanding of whether a tool call is appropriate or dangerous. If an LLM produces an instruction that resembles a tool invocation, that does not mean the model has "acted" - it means the system around the model has chosen to treat that output as executable.

Why Consumer AI Exposes More Surface Area

Consumer focused AI products typically prioritise convenience and flexibility. As a result, they often expose a wide range of tools by default:

  • Email and messaging
  • File access and sharing
  • Web browsing
  • Calendars and task systems
  • Third-party plugins

The more tools that are available, the larger the attack surface becomes. If untrusted content is introduced into the model’s context , via retrieved documents, pasted text, or user input, the model may generate outputs that attempt to invoke tools in unexpected ways. Several well-known examples originate in environments where broad tool access is enabled by design.

Enterprise AI Has Different Requirements

Enterprise AI systems on the other hand should be built for specific outcomes, bounded workflows, and defined responsibility. In an enterprise context:

  • Tools should be enabled only when required
  • Tool schemas should be explicit and validated
  • Execution paths should be constrained
  • Behaviour should be observable and auditable

For example, if an AI workflow is performing structured data extraction or classification, there is typically no reason for it to have access to email, file-sharing, or outbound communication tools. Those capabilities should not exist in that execution context.

Security Is a Platform Property, Not a Model Feature

There is no such thing as a "secure LLM" in isolation. Security emerges from system design:

  • What tools are available
  • How inputs are validated
  • How outputs are interpreted
  • What actions are permitted
  • What is logged and reviewed

An LLM can suggest an action, but the platform decides whether that action is allowed, how it is executed, and whether it is rejected.

How Zeaware Avalon Approaches Tool Control

Zeaware Avalon is designed on the assumption that capability must be explicit. From an engineering perspective, this means:

  • Tools are not globally available enabled
  • Each workflow, and each task in the work, explicitly declares which tools it may use
  • Tool inputs are validated before execution
  • Tool execution is controlled by the platform, not the model
  • Outputs and decisions are captured for audit and review

In many enterprise scenarios, the safest configuration is one with no tools enabled at all, beyond retrieval and reasoning. When tools are required, they are treated as governed execution steps - not conveniences the model can freely explore.

LLM Suggestion vs Platform-Controlled Execution

User / Retrieved Content LLM Generates text or suggestions Enterprise AI Platform · Tool allow-list · Schema validation · Execution rules · Audit Approved Tool (Scoped Action)

Addressing Common Objections

"Can’t an LLM still generate malicious tool calls even with limited tools?"

Yes. Tool restriction stops models from accessing tools out of scope for the assigned task.  But this alone is not a complete solution. Limiting tools reduces surface area, but validation and enforcement prevent misuse within that surface area.

"What about prompt injection through retrieved content?"

Retrieved content should be treated as untrusted. It should inform reasoning, not expand capability. Tool availability and execution authority must remain independent of retrieved data.

"Doesn’t the model still need to behave correctly?"

Models are probabilistic by nature. Enterprise systems should assume models may produce unexpected outputs, and ensure those outputs cannot trigger unauthorised actions.  Guardrails are used to validate and revise, redo and fail unexpected results.

"What about incorrect reasoning or misleading outputs?"

That is a separate class of risk. Tool control does not eliminate reasoning errors or hallucinations - those require different mitigations such as evaluation, review, and governance processes.

A Balanced View of Risk

AI systems introduce new considerations, but they do not invalidate decades of security practice. There will always be risks to manage: untrusted inputs, misconfiguration, over-permissioned execution, and insufficient monitoring. The goal is to understand where risk lives, reduce exposure through design, and make behaviour observable.

— Zeaware Engineering Team

]]> Recent discussion around tool exfiltration and indirect prompt injection attacks in Generative AI systems has raised valid concerns - particularly where platforms unexpectedly invoke tools or actions as a result of untrusted input. These concerns are worth taking seriously. But they are often framed in a way that blurs the distinction between the model, the application, and the execution platform.

This post clarifies that distinction, explains where risk actually lives, and outlines how enterprise AI systems can be designed to limit exposure through explicit control.

The LLM Is Just a Model

At its core, a large language model (LLM) does one thing: it takes an input and generates an output. That output may be plain text for a human, or it may be a structured response that an application interprets as a suggestion to call a tool.

The model itself has no execution capability, no awareness of trust boundaries, and no understanding of whether a tool call is appropriate or dangerous. If an LLM produces an instruction that resembles a tool invocation, that does not mean the model has "acted" - it means the system around the model has chosen to treat that output as executable.

Why Consumer AI Exposes More Surface Area

Consumer focused AI products typically prioritise convenience and flexibility. As a result, they often expose a wide range of tools by default:

  • Email and messaging
  • File access and sharing
  • Web browsing
  • Calendars and task systems
  • Third-party plugins

The more tools that are available, the larger the attack surface becomes. If untrusted content is introduced into the model’s context , via retrieved documents, pasted text, or user input, the model may generate outputs that attempt to invoke tools in unexpected ways. Several well-known examples originate in environments where broad tool access is enabled by design.

Enterprise AI Has Different Requirements

Enterprise AI systems on the other hand should be built for specific outcomes, bounded workflows, and defined responsibility. In an enterprise context:

  • Tools should be enabled only when required
  • Tool schemas should be explicit and validated
  • Execution paths should be constrained
  • Behaviour should be observable and auditable

For example, if an AI workflow is performing structured data extraction or classification, there is typically no reason for it to have access to email, file-sharing, or outbound communication tools. Those capabilities should not exist in that execution context.

Security Is a Platform Property, Not a Model Feature

There is no such thing as a "secure LLM" in isolation. Security emerges from system design:

  • What tools are available
  • How inputs are validated
  • How outputs are interpreted
  • What actions are permitted
  • What is logged and reviewed

An LLM can suggest an action, but the platform decides whether that action is allowed, how it is executed, and whether it is rejected.

How Zeaware Avalon Approaches Tool Control

Zeaware Avalon is designed on the assumption that capability must be explicit. From an engineering perspective, this means:

  • Tools are not globally available enabled
  • Each workflow, and each task in the work, explicitly declares which tools it may use
  • Tool inputs are validated before execution
  • Tool execution is controlled by the platform, not the model
  • Outputs and decisions are captured for audit and review

In many enterprise scenarios, the safest configuration is one with no tools enabled at all, beyond retrieval and reasoning. When tools are required, they are treated as governed execution steps - not conveniences the model can freely explore.

LLM Suggestion vs Platform-Controlled Execution

User / Retrieved Content LLM Generates text or suggestions Enterprise AI Platform · Tool allow-list · Schema validation · Execution rules · Audit Approved Tool (Scoped Action)

Addressing Common Objections

"Can’t an LLM still generate malicious tool calls even with limited tools?"

Yes. Tool restriction stops models from accessing tools out of scope for the assigned task.  But this alone is not a complete solution. Limiting tools reduces surface area, but validation and enforcement prevent misuse within that surface area.

"What about prompt injection through retrieved content?"

Retrieved content should be treated as untrusted. It should inform reasoning, not expand capability. Tool availability and execution authority must remain independent of retrieved data.

"Doesn’t the model still need to behave correctly?"

Models are probabilistic by nature. Enterprise systems should assume models may produce unexpected outputs, and ensure those outputs cannot trigger unauthorised actions.  Guardrails are used to validate and revise, redo and fail unexpected results.

"What about incorrect reasoning or misleading outputs?"

That is a separate class of risk. Tool control does not eliminate reasoning errors or hallucinations - those require different mitigations such as evaluation, review, and governance processes.

A Balanced View of Risk

AI systems introduce new considerations, but they do not invalidate decades of security practice. There will always be risks to manage: untrusted inputs, misconfiguration, over-permissioned execution, and insufficient monitoring. The goal is to understand where risk lives, reduce exposure through design, and make behaviour observable.

— Zeaware Engineering Team

]]>> Objectives for 2026 https://live.cms.zeaware.net/zeaware/blog/2026_year_ahead Fri, 02 Jan 2026 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/2026_year_ahead #zeaware-2026-trust-ai .zw-eyebrow{ font-size: 13px; letter-spacing: .08em; text-transform: uppercase; color: #6a6a6a; margin: 0 0 8px 0; } #zeaware-2026-trust-ai .zw-lede{ font-size: 18px; color: #2b2b2b; margin: 0 0 18px 0; } #zeaware-2026-trust-ai ul{ margin: 8px 0 14px 18px; padding: 0; } #zeaware-2026-trust-ai li{ margin: 6px 0; } #zeaware-2026-trust-ai .zw-grid{ display: grid; grid-template-columns: 1fr; gap: 14px; margin: 14px 0 8px 0; } @media (min-width: 860px){ #zeaware-2026-trust-ai .zw-grid{ grid-template-columns: 1fr 1fr; } } #zeaware-2026-trust-ai .zw-card{ border: 1px solid #ececec; border-radius: 14px; padding: 14px 14px 12px 14px; background: #fff; box-shadow: 0 6px 18px rgba(0,0,0,0.04); } #zeaware-2026-trust-ai .zw-card h4{ margin: 0 0 6px 0; font-size: 16px; } #zeaware-2026-trust-ai .zw-card p{ margin: 0 0 8px 0; color: #2f2f2f; font-size: 14px; } #zeaware-2026-trust-ai .zw-tagrow{ display: flex; flex-wrap: wrap; gap: 8px; margin-top: 10px; } #zeaware-2026-trust-ai .zw-tag{ font-size: 12px; padding: 5px 9px; border-radius: 999px; border: 1px solid #e9e9e9; background: #fafafa; color: #3a3a3a; white-space: nowrap; }

2026 - The Year Trust Becomes the Real AI Advantage

As we step into 2026, the AI conversation is shifting in a healthy way.  Less of “can it do it?”, more of “can we trust it, and can we prove it?”. That shift matters because agentic systems are moving from demos into production and when AI can initiate change in the real world, trust stops being a principle and becomes a commercial requirement.

Our core belief for 2026: If you can’t explain what happened, you can’t scale it. Trust is not a slogan, it is an engineering discipline.

What will define AI success in 2026

  • ROI pressure meets reality - experimentation is giving way to dependable, repeatable productivity gains and measurable outcomes.
  • Agents go mainstream - and break in new ways - reliability, traceability, and “safe autonomy” become the hard problems.
  • Trust becomes the currency - governance and control shift from “compliance work” to the enabler of scale.

Zeaware’s top 5 objectives for 2026

1) Make “Trusted AI” practical

Turn trust into something you can build, configure, audit, and improve, not something you hope for. In Zeaware Avalon, this means governance-by-design, policy-aware execution, and end-to-end traceability.

Governance-by-design Audit trails Policy-aware execution

2) Operationalise agentic workflows safely

Move beyond “chat” into real workflows while keeping humans in control where it matters. The goal is not autonomy for its own sake,-it is about safe and scalable leverage.

Human-in-the-loop Scoped automation Workflow patterns

3) Raise AI reliability with measurable quality

In 2026, “it seemed right” won’t pass. We’re investing in evaluation harnesses, regression tests for workflows, and operational monitoring to detect drift and repeated failure patterns.

Evaluation Regression testing Monitoring

4) Help customers defend trust, not just create content

As threats like impersonation, scams, and agent misuse rise, organisations will invest heavily in defence. Zeaware’s role is help provide controlled tool access, identity and permissions, and auditability.

Identity Least privilege Auditability

5) Build for enterprise-grade realities

Customers need clarity on data boundaries, deployment models, and operational control. We’ll keep investing in multi-tenant governance and customer-hosted options suitable for regulated environments.

Data boundaries Customer-hosted Enterprise security

What this means in practice

Start with real workflows and measurable outcomes. Design trust, control, and auditability from day one. Deploy agents that can act - but only inside explicit boundaries. Build the muscle of evaluation and continuous improvement.

Real workflows Boundaries Continuous improvement

Partner with Zeaware in 2026

Our focus in 2026 is not just on building trusted AI systems, but on helping organisations grow with AI, safely, sustainably, and at enterprise scale.

That means meeting customers where they are today, whether they are moving beyond early pilots or looking to scale existing AI capabilities across teams, functions, and geographies. We work closely with customers to prioritise the right use cases, design agentic workflows that deliver real operational value, and embed trust, governance, and accountability from day one so growth does not come at the cost of control.

We also recognise that scaling AI is as much a delivery challenge as it is a technology one. That is why Zeaware works with a growing network of delivery and domain partners who bring deep industry expertise, implementation capability, and change management experience. Together, we provide a scalable delivery model that allows organisations to move faster, deploy confidently, and extend AI across the business without reinventing the wheel each time.

Zeaware Avalon sits at the centre of this approach - providing a governed, extensible platform that enables customers and partners to build, operate, and continuously improve AI solutions in production, not just experiment in isolation.

If you’re planning AI for 2026 and want to move from experimentation to trusted production systems, systems that your people can rely on, your leadership can stand behind, and your organisation can scale with confidence, we’d love to collaborate.

]]> #zeaware-2026-trust-ai .zw-eyebrow{ font-size: 13px; letter-spacing: .08em; text-transform: uppercase; color: #6a6a6a; margin: 0 0 8px 0; } #zeaware-2026-trust-ai .zw-lede{ font-size: 18px; color: #2b2b2b; margin: 0 0 18px 0; } #zeaware-2026-trust-ai ul{ margin: 8px 0 14px 18px; padding: 0; } #zeaware-2026-trust-ai li{ margin: 6px 0; } #zeaware-2026-trust-ai .zw-grid{ display: grid; grid-template-columns: 1fr; gap: 14px; margin: 14px 0 8px 0; } @media (min-width: 860px){ #zeaware-2026-trust-ai .zw-grid{ grid-template-columns: 1fr 1fr; } } #zeaware-2026-trust-ai .zw-card{ border: 1px solid #ececec; border-radius: 14px; padding: 14px 14px 12px 14px; background: #fff; box-shadow: 0 6px 18px rgba(0,0,0,0.04); } #zeaware-2026-trust-ai .zw-card h4{ margin: 0 0 6px 0; font-size: 16px; } #zeaware-2026-trust-ai .zw-card p{ margin: 0 0 8px 0; color: #2f2f2f; font-size: 14px; } #zeaware-2026-trust-ai .zw-tagrow{ display: flex; flex-wrap: wrap; gap: 8px; margin-top: 10px; } #zeaware-2026-trust-ai .zw-tag{ font-size: 12px; padding: 5px 9px; border-radius: 999px; border: 1px solid #e9e9e9; background: #fafafa; color: #3a3a3a; white-space: nowrap; }

2026 - The Year Trust Becomes the Real AI Advantage

As we step into 2026, the AI conversation is shifting in a healthy way.  Less of “can it do it?”, more of “can we trust it, and can we prove it?”. That shift matters because agentic systems are moving from demos into production and when AI can initiate change in the real world, trust stops being a principle and becomes a commercial requirement.

Our core belief for 2026: If you can’t explain what happened, you can’t scale it. Trust is not a slogan, it is an engineering discipline.

What will define AI success in 2026

  • ROI pressure meets reality - experimentation is giving way to dependable, repeatable productivity gains and measurable outcomes.
  • Agents go mainstream - and break in new ways - reliability, traceability, and “safe autonomy” become the hard problems.
  • Trust becomes the currency - governance and control shift from “compliance work” to the enabler of scale.

Zeaware’s top 5 objectives for 2026

1) Make “Trusted AI” practical

Turn trust into something you can build, configure, audit, and improve, not something you hope for. In Zeaware Avalon, this means governance-by-design, policy-aware execution, and end-to-end traceability.

Governance-by-design Audit trails Policy-aware execution

2) Operationalise agentic workflows safely

Move beyond “chat” into real workflows while keeping humans in control where it matters. The goal is not autonomy for its own sake,-it is about safe and scalable leverage.

Human-in-the-loop Scoped automation Workflow patterns

3) Raise AI reliability with measurable quality

In 2026, “it seemed right” won’t pass. We’re investing in evaluation harnesses, regression tests for workflows, and operational monitoring to detect drift and repeated failure patterns.

Evaluation Regression testing Monitoring

4) Help customers defend trust, not just create content

As threats like impersonation, scams, and agent misuse rise, organisations will invest heavily in defence. Zeaware’s role is help provide controlled tool access, identity and permissions, and auditability.

Identity Least privilege Auditability

5) Build for enterprise-grade realities

Customers need clarity on data boundaries, deployment models, and operational control. We’ll keep investing in multi-tenant governance and customer-hosted options suitable for regulated environments.

Data boundaries Customer-hosted Enterprise security

What this means in practice

Start with real workflows and measurable outcomes. Design trust, control, and auditability from day one. Deploy agents that can act - but only inside explicit boundaries. Build the muscle of evaluation and continuous improvement.

Real workflows Boundaries Continuous improvement

Partner with Zeaware in 2026

Our focus in 2026 is not just on building trusted AI systems, but on helping organisations grow with AI, safely, sustainably, and at enterprise scale.

That means meeting customers where they are today, whether they are moving beyond early pilots or looking to scale existing AI capabilities across teams, functions, and geographies. We work closely with customers to prioritise the right use cases, design agentic workflows that deliver real operational value, and embed trust, governance, and accountability from day one so growth does not come at the cost of control.

We also recognise that scaling AI is as much a delivery challenge as it is a technology one. That is why Zeaware works with a growing network of delivery and domain partners who bring deep industry expertise, implementation capability, and change management experience. Together, we provide a scalable delivery model that allows organisations to move faster, deploy confidently, and extend AI across the business without reinventing the wheel each time.

Zeaware Avalon sits at the centre of this approach - providing a governed, extensible platform that enables customers and partners to build, operate, and continuously improve AI solutions in production, not just experiment in isolation.

If you’re planning AI for 2026 and want to move from experimentation to trusted production systems, systems that your people can rely on, your leadership can stand behind, and your organisation can scale with confidence, we’d love to collaborate.

]]>> Free eBook on Agentic Governance https://live.cms.zeaware.net/zeaware/blog/free_ebook_agentic_ai_governance Wed, 17 Dec 2025 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/free_ebook_agentic_ai_governance

A Holiday Gift for AI Leaders - Free eBook on Agentic Governance

As organisations accelerate the adoption of AI and agentic systems, one question keeps surfacing: How do we move fast with AI - without losing control?

To help answer that question, Zeaware is pleased to offer our community a free download of Agentic Governance: Making Intelligence Trustworthy by Jesper Lowgren.

@{#imageembed("free_ebook_agentic_ai_governance_image")}

Why Agentic Governance Matters Now

Agentic AI represents a step change in how systems operate. These are not just models responding to prompts - they are systems that can plan, decide, act, and collaborate across tools, data, and workflows.

That power creates opportunity, but it also raises new challenges:

  • How do we ensure AI actions remain aligned with organisational policy?
  • How do we embed accountability when systems are autonomous?
  • How do we govern outcomes, not just access?

This eBook explores these questions through a practical governance lens, focusing on design-time and runtime controls - not just policy documents or after-the-fact oversight.

What You’ll Learn

In Agentic Governance, Jesper Lowgren draws on deep experience in AI systems design to explain:

  • Why traditional AI governance models are no longer sufficient
  • How governance can be embedded into agent architecture itself
  • The role of policy-as-data, constraints, and feedback loops
  • How organisations can scale AI safely without stifling innovation

The book is platform-agnostic and written for leaders, architects, and practitioners who want to move beyond theory into real-world application.

A Thank You to the Author

We would like to sincerely thank Jesper Lowgren for allowing Zeaware to make this eBook freely available to our community - particularly at a time when many teams are reflecting, planning, and setting direction for the year ahead.

Jesper’s work aligns closely with Zeaware’s own philosophy: governance should enable AI adoption, not block it - and trust must be designed into systems from day one.

Just in Time for Planning Season

Whether you are exploring your first agent-based AI use cases, scaling AI across business units, or preparing governance frameworks for 2026 and beyond, this eBook provides a thoughtful, practical foundation to support those conversations.

Get the free eBook

]]>

A Holiday Gift for AI Leaders - Free eBook on Agentic Governance

As organisations accelerate the adoption of AI and agentic systems, one question keeps surfacing: How do we move fast with AI - without losing control?

To help answer that question, Zeaware is pleased to offer our community a free download of Agentic Governance: Making Intelligence Trustworthy by Jesper Lowgren.

@{#imageembed("free_ebook_agentic_ai_governance_image")}

Why Agentic Governance Matters Now

Agentic AI represents a step change in how systems operate. These are not just models responding to prompts - they are systems that can plan, decide, act, and collaborate across tools, data, and workflows.

That power creates opportunity, but it also raises new challenges:

  • How do we ensure AI actions remain aligned with organisational policy?
  • How do we embed accountability when systems are autonomous?
  • How do we govern outcomes, not just access?

This eBook explores these questions through a practical governance lens, focusing on design-time and runtime controls - not just policy documents or after-the-fact oversight.

What You’ll Learn

In Agentic Governance, Jesper Lowgren draws on deep experience in AI systems design to explain:

  • Why traditional AI governance models are no longer sufficient
  • How governance can be embedded into agent architecture itself
  • The role of policy-as-data, constraints, and feedback loops
  • How organisations can scale AI safely without stifling innovation

The book is platform-agnostic and written for leaders, architects, and practitioners who want to move beyond theory into real-world application.

A Thank You to the Author

We would like to sincerely thank Jesper Lowgren for allowing Zeaware to make this eBook freely available to our community - particularly at a time when many teams are reflecting, planning, and setting direction for the year ahead.

Jesper’s work aligns closely with Zeaware’s own philosophy: governance should enable AI adoption, not block it - and trust must be designed into systems from day one.

Just in Time for Planning Season

Whether you are exploring your first agent-based AI use cases, scaling AI across business units, or preparing governance frameworks for 2026 and beyond, this eBook provides a thoughtful, practical foundation to support those conversations.

Get the free eBook

]]>> GPT-5.2 now supported in Zeaware Avalon https://live.cms.zeaware.net/zeaware/blog/avalon_openai_gpt_5_2_support Fri, 12 Dec 2025 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/avalon_openai_gpt_5_2_support We are pleased to announce that Zeaware Avalon now supports OpenAI’s latest flagship model, GPT-5.2. GPT-5.2 is designed for professional knowledge work and long-running agents, with major improvements in reasoning, long-context understanding, tool use and multimodal capabilities.

What’s new in GPT-5.2

  • Enhanced reasoning and professional task execution – GPT-5.2 is optimised for complex, well-specified knowledge work such as building spreadsheets, presentations, technical documentation and production-grade code.
  • Stronger long-context and multimodal performance – Better handling of extended inputs and mixed modalities (text, code and images) helps maintain fidelity across complex workflows and larger document sets.
  • Improved agentic tool-calling – More reliable tool use and planning for agents that need to call APIs, work with structured data or orchestrate multi-step business processes end-to-end.
  • Broader availability – GPT-5.2 is available in multiple variants for different latency, cost and depth-of-thinking profiles, making it easier to match models to real-world use cases.

Benefits for Zeaware Avalon customers

  • Smarter agents and workflows – Power demanding workloads such as contract and policy analysis, research assistance, technical support, data exploration and decision support with a frontier-grade reasoning model.
  • More reliable long-running interactions – Use GPT-5.2 in agents that operate over longer sessions, larger knowledge bases or multi-step workflows while maintaining consistency and context.
  • Richer automation opportunities – Combine  Zeaware Avalon’s governance-by-design framework with GPT-5.2’s improved planning and tool-use to automate more of the “last mile” in complex processes, while keeping humans in control.
  • Flexible model choice – Configure GPT-5.2 alongside other supported models in Zeaware Avalon, choosing the right engine per agent, per workflow or per tenant based on cost, performance and risk profile.

Getting started

GPT-5.2 can be added via the Service settings in your Zeaware Avalon environment. Once enabled, you can:

  • Re-run existing agent and workflow tests using GPT-5.2 to compare output quality, robustness and latency.
  • Introduce GPT-5.2 into complex, multi-step workflows that benefit from stronger reasoning and tool-use.
  • Review workflows that currently require manual oversight to see where GPT-5.2 can safely reduce orchestration effort.

What’s next

For a demonstration of GPT-5.2 within Zeaware Avalon or guidance on upgrading your agents and workflows, please contact the Zeaware team.

]]> We are pleased to announce that Zeaware Avalon now supports OpenAI’s latest flagship model, GPT-5.2. GPT-5.2 is designed for professional knowledge work and long-running agents, with major improvements in reasoning, long-context understanding, tool use and multimodal capabilities.

What’s new in GPT-5.2

  • Enhanced reasoning and professional task execution – GPT-5.2 is optimised for complex, well-specified knowledge work such as building spreadsheets, presentations, technical documentation and production-grade code.
  • Stronger long-context and multimodal performance – Better handling of extended inputs and mixed modalities (text, code and images) helps maintain fidelity across complex workflows and larger document sets.
  • Improved agentic tool-calling – More reliable tool use and planning for agents that need to call APIs, work with structured data or orchestrate multi-step business processes end-to-end.
  • Broader availability – GPT-5.2 is available in multiple variants for different latency, cost and depth-of-thinking profiles, making it easier to match models to real-world use cases.

Benefits for Zeaware Avalon customers

  • Smarter agents and workflows – Power demanding workloads such as contract and policy analysis, research assistance, technical support, data exploration and decision support with a frontier-grade reasoning model.
  • More reliable long-running interactions – Use GPT-5.2 in agents that operate over longer sessions, larger knowledge bases or multi-step workflows while maintaining consistency and context.
  • Richer automation opportunities – Combine  Zeaware Avalon’s governance-by-design framework with GPT-5.2’s improved planning and tool-use to automate more of the “last mile” in complex processes, while keeping humans in control.
  • Flexible model choice – Configure GPT-5.2 alongside other supported models in Zeaware Avalon, choosing the right engine per agent, per workflow or per tenant based on cost, performance and risk profile.

Getting started

GPT-5.2 can be added via the Service settings in your Zeaware Avalon environment. Once enabled, you can:

  • Re-run existing agent and workflow tests using GPT-5.2 to compare output quality, robustness and latency.
  • Introduce GPT-5.2 into complex, multi-step workflows that benefit from stronger reasoning and tool-use.
  • Review workflows that currently require manual oversight to see where GPT-5.2 can safely reduce orchestration effort.

What’s next

For a demonstration of GPT-5.2 within Zeaware Avalon or guidance on upgrading your agents and workflows, please contact the Zeaware team.

]]>> Tenant-Level Firewalls Now Available in Zeaware Avalon https://live.cms.zeaware.net/zeaware/blog/zeaware_avalon_tenant_firewall Sun, 30 Nov 2025 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/zeaware_avalon_tenant_firewall Zeaware Avalon now includes tenant-level firewalls, giving you direct control over which networks are allowed to access your environment. This feature lets you restrict access to your tenant using CIDR-based IP allow lists, providing an additional layer of protection for your AI agents, data, and workflows.

When the firewall is enabled, Zeaware Avalon switches to a default-deny model. Only the IP ranges you explicitly approve will be permitted, mirroring the defence-in-depth approach used by leading cloud platforms. This helps ensure your environment remains secure, predictable, and aligned with your organisation’s policies.

  • Choose between Allow access from all networks or Allow access from selected networks
  • Add one or more CIDR IP ranges to define who can access your tenant
  • Built-in guardrails help reduce the risk of accidental lockouts
  • All configuration changes are tracked for auditing and compliance

This enhancement is part of our ongoing commitment to delivering enterprise-grade security and governance within Avalon. As organisations scale their use of AI agents and automations, strong access controls are essential for protecting data and ensuring reliable operation.

@{#imageviewer("zeaware_avalon_tenant_firewall_example")}

Tenant firewalls are now available in the Network section of the Zeaware Avalon console. If you’d like guidance on recommended configurations or help getting started, our support team is ready to assist.

]]> Zeaware Avalon now includes tenant-level firewalls, giving you direct control over which networks are allowed to access your environment. This feature lets you restrict access to your tenant using CIDR-based IP allow lists, providing an additional layer of protection for your AI agents, data, and workflows.

When the firewall is enabled, Zeaware Avalon switches to a default-deny model. Only the IP ranges you explicitly approve will be permitted, mirroring the defence-in-depth approach used by leading cloud platforms. This helps ensure your environment remains secure, predictable, and aligned with your organisation’s policies.

  • Choose between Allow access from all networks or Allow access from selected networks
  • Add one or more CIDR IP ranges to define who can access your tenant
  • Built-in guardrails help reduce the risk of accidental lockouts
  • All configuration changes are tracked for auditing and compliance

This enhancement is part of our ongoing commitment to delivering enterprise-grade security and governance within Avalon. As organisations scale their use of AI agents and automations, strong access controls are essential for protecting data and ensuring reliable operation.

@{#imageviewer("zeaware_avalon_tenant_firewall_example")}

Tenant firewalls are now available in the Network section of the Zeaware Avalon console. If you’d like guidance on recommended configurations or help getting started, our support team is ready to assist.

]]>> How Zeaware Avalon Optimises Content for Agents https://live.cms.zeaware.net/zeaware/blog/how_zeaware_avalon_optimises_content_for_agents Sat, 29 Nov 2025 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/how_zeaware_avalon_optimises_content_for_agents Integrated Chunking, Vector Indexing, and Hybrid Search

Enterprise AI systems rarely fail because just because of model capability, instead one of the key reasons they may fail is because the content pipeline feeding those models is incomplete, inconsistent, or poorly controlled.

At Zeaware, we have seen this repeatedly across customer environments: documents that look simple on paper behave unpredictably when converted, split, embedded, and retrieved. Organisations often try to solve this by using enterprise search tools, only to discover later that they have created a fragmented architecture full of hidden cost, latency, and maintenance overhead.

Zeaware Avalon was designed to help improve this.

Rather than treating content preparation, chunking, embedding and retrieval as an external concern, Zeaware Avalon provides an integrated set of capabilities that optimise content for agent reasoning, while still giving customers the freedom to plug in large-scale search services when needed.

This hybrid approach delivers the best of both worlds: performance, control and efficiency internally, and elastic scale externally.

Content-Aware Splitting Designed for Agent Reasoning

Chunking is deceptively complex. Most systems use naïve methods - fixed-size tokens or simple paragraph splits - which leads to:

  • Missing or truncated context
  • Hallucinated linkages between unrelated sections
  • High token usage
  • Reduced answer accuracy in multi-step agent workflows

Zeaware Avalon’s content-aware splitter takes a different approach.

It looks at structural cues (headings, semantic boundaries, topic shifts, embedded tables, lists and captions) and optimises splits so that each section:

  • Preserves its topic integrity
  • Includes essential upstream context
  • Produces high-quality embeddings
  • Enables more deterministic retrieval for agents

This is particularly important for multi-agent workflows. When an agent summarises, extracts, cross-checks, or synthesises content, chunk quality determines reasoning quality.

Zeaware Avalon handles this automatically. In Auto mode there is nothing for customers to configure, tune or maintain however fine grained control can be invoke when required.

Integrated Vector Indexing for Agent-Local Content

Many agents operate on small, specialised content pools:

  • A product spec booklet for an onboarding agent
  • A single policy document for an HR Q&A agent
  • A structured dataset for a calculation or classification agent
  • A narrow knowledge bundle for an audit or compliance flow

Provisioning an external vector index for each of these may be:

  • Unnecessary
  • Expensive
  • Operationally messy
  • Slow for continuous development and experimentation

Zeaware Avalon provides its own in-built vector index optimised for agent-local use cases. This gives each agent fast, high-quality semantic retrieval without requiring external services.

This delivers three important advantages:

Zero infrastructure overhead

No requirement for external search services (when small content sets), no provisioning limits.

High speed, low latency

Local vector stores are optimised for high speed lookups, perfect for multi-step agent workflows where every millisecond compounds.

Better governance-by-design

Zeaware Avalon knows exactly which content an agent has access to and enforces retrieval boundaries without relying on external system configuration.

External Index Integration for Large-Scale Content Pools

Of course, many enterprise use cases require large-scale document pools, tens of thousands of files, or continuously updated corporate knowledge bases.

Zeaware Avalon does not replace that. instead we extend using these services.

Agents can seamlessly reference external search indexes when:

  • Document volumes exceed internal storage targets
  • Content is updated regularly across the organisation
  • Search must span multiple systems (SharePoint, websites, file repositories)
  • Teams want to reuse existing search infrastructure

Zeaware Avalon handles the complexity:

  • Schema mapping
  • Index connection
  • Embedding model alignment
  • Result scoring
  • Governance and retrieval limits

This gives customers the flexibility of external search with the consistency and safety of a Zeaware Avalon-managed pipeline.

Why This Integrated Design Matters

Teams building enterprise AI solutions often discover hidden costs later:

  • Multiple external indexes per agent or project
  • Duplicated document ingestion pipelines
  • Fragmented governance rules
  • Inconsistent splitting logic
  • High search costs
  • Retrieval drift as different indexes diverge
  • Non-deterministic responses because chunking and context rules vary

Zeaware Avalon helps reduce this by providing one unified content optimisation framework, improving predictable behaviour and consistent governance, regardless of whether the agent uses:

  • Internal vector indexing
  • External search integration
  • Or both

This reduces operational burden dramatically and ensures that AI agents behave reliably, even as organisations scale their usage.

Delivering Better Accuracy, Lower Cost, and Stronger Governance

Zeaware Avalon’s integrated approach gives customers three strategic benefits:

Accurate and deterministic agent reasoning

Because Zeaware Avalon controls chunking, embedding, and retrieval, agents receive more consistent and contextually correct information.

Lower operational overhead

Most agent content can stay internal. Large-scale search can be external. This hybrid model keeps infrastructure light and costs predictable.

Strong governance and transparency

Zeaware Avalon knows what each agent can see and controls how it retrieves it - which is essential for enterprise governance, auditability, and risk management.

Zeaware Avalon: A Platform Built for Real-World Agent Performance

Content is the fuel of enterprise AI. If the preparation pipeline is inconsistent, expensive, or fragmented, agent performance will never be reliable.

Zeaware Avalon solves this with a unified, content-aware design:

  • Intelligent chunking
  • Robust local vector indexes
  • Seamless integration with enterprise search
  • Consistent metadata and governance controls

This allows organisations to move fast, build reliably, and scale without multiplying infrastructure or cost. It is about having the right content, retrieved under the right governance, at the right time.

Zeaware Engineering

]]> Integrated Chunking, Vector Indexing, and Hybrid Search

Enterprise AI systems rarely fail because just because of model capability, instead one of the key reasons they may fail is because the content pipeline feeding those models is incomplete, inconsistent, or poorly controlled.

At Zeaware, we have seen this repeatedly across customer environments: documents that look simple on paper behave unpredictably when converted, split, embedded, and retrieved. Organisations often try to solve this by using enterprise search tools, only to discover later that they have created a fragmented architecture full of hidden cost, latency, and maintenance overhead.

Zeaware Avalon was designed to help improve this.

Rather than treating content preparation, chunking, embedding and retrieval as an external concern, Zeaware Avalon provides an integrated set of capabilities that optimise content for agent reasoning, while still giving customers the freedom to plug in large-scale search services when needed.

This hybrid approach delivers the best of both worlds: performance, control and efficiency internally, and elastic scale externally.

Content-Aware Splitting Designed for Agent Reasoning

Chunking is deceptively complex. Most systems use naïve methods - fixed-size tokens or simple paragraph splits - which leads to:

  • Missing or truncated context
  • Hallucinated linkages between unrelated sections
  • High token usage
  • Reduced answer accuracy in multi-step agent workflows

Zeaware Avalon’s content-aware splitter takes a different approach.

It looks at structural cues (headings, semantic boundaries, topic shifts, embedded tables, lists and captions) and optimises splits so that each section:

  • Preserves its topic integrity
  • Includes essential upstream context
  • Produces high-quality embeddings
  • Enables more deterministic retrieval for agents

This is particularly important for multi-agent workflows. When an agent summarises, extracts, cross-checks, or synthesises content, chunk quality determines reasoning quality.

Zeaware Avalon handles this automatically. In Auto mode there is nothing for customers to configure, tune or maintain however fine grained control can be invoke when required.

Integrated Vector Indexing for Agent-Local Content

Many agents operate on small, specialised content pools:

  • A product spec booklet for an onboarding agent
  • A single policy document for an HR Q&A agent
  • A structured dataset for a calculation or classification agent
  • A narrow knowledge bundle for an audit or compliance flow

Provisioning an external vector index for each of these may be:

  • Unnecessary
  • Expensive
  • Operationally messy
  • Slow for continuous development and experimentation

Zeaware Avalon provides its own in-built vector index optimised for agent-local use cases. This gives each agent fast, high-quality semantic retrieval without requiring external services.

This delivers three important advantages:

Zero infrastructure overhead

No requirement for external search services (when small content sets), no provisioning limits.

High speed, low latency

Local vector stores are optimised for high speed lookups, perfect for multi-step agent workflows where every millisecond compounds.

Better governance-by-design

Zeaware Avalon knows exactly which content an agent has access to and enforces retrieval boundaries without relying on external system configuration.

External Index Integration for Large-Scale Content Pools

Of course, many enterprise use cases require large-scale document pools, tens of thousands of files, or continuously updated corporate knowledge bases.

Zeaware Avalon does not replace that. instead we extend using these services.

Agents can seamlessly reference external search indexes when:

  • Document volumes exceed internal storage targets
  • Content is updated regularly across the organisation
  • Search must span multiple systems (SharePoint, websites, file repositories)
  • Teams want to reuse existing search infrastructure

Zeaware Avalon handles the complexity:

  • Schema mapping
  • Index connection
  • Embedding model alignment
  • Result scoring
  • Governance and retrieval limits

This gives customers the flexibility of external search with the consistency and safety of a Zeaware Avalon-managed pipeline.

Why This Integrated Design Matters

Teams building enterprise AI solutions often discover hidden costs later:

  • Multiple external indexes per agent or project
  • Duplicated document ingestion pipelines
  • Fragmented governance rules
  • Inconsistent splitting logic
  • High search costs
  • Retrieval drift as different indexes diverge
  • Non-deterministic responses because chunking and context rules vary

Zeaware Avalon helps reduce this by providing one unified content optimisation framework, improving predictable behaviour and consistent governance, regardless of whether the agent uses:

  • Internal vector indexing
  • External search integration
  • Or both

This reduces operational burden dramatically and ensures that AI agents behave reliably, even as organisations scale their usage.

Delivering Better Accuracy, Lower Cost, and Stronger Governance

Zeaware Avalon’s integrated approach gives customers three strategic benefits:

Accurate and deterministic agent reasoning

Because Zeaware Avalon controls chunking, embedding, and retrieval, agents receive more consistent and contextually correct information.

Lower operational overhead

Most agent content can stay internal. Large-scale search can be external. This hybrid model keeps infrastructure light and costs predictable.

Strong governance and transparency

Zeaware Avalon knows what each agent can see and controls how it retrieves it - which is essential for enterprise governance, auditability, and risk management.

Zeaware Avalon: A Platform Built for Real-World Agent Performance

Content is the fuel of enterprise AI. If the preparation pipeline is inconsistent, expensive, or fragmented, agent performance will never be reliable.

Zeaware Avalon solves this with a unified, content-aware design:

  • Intelligent chunking
  • Robust local vector indexes
  • Seamless integration with enterprise search
  • Consistent metadata and governance controls

This allows organisations to move fast, build reliably, and scale without multiplying infrastructure or cost. It is about having the right content, retrieved under the right governance, at the right time.

Zeaware Engineering

]]>> Claude Opus 4.5 now supported in Zeaware Avalon https://live.cms.zeaware.net/zeaware/blog/avalon_claude_opus_4_5_support Tue, 25 Nov 2025 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/avalon_claude_opus_4_5_support We are pleased to announce that Zeaware Avalon now supports Anthropic’s latest flagship model, Claude Opus 4.5. Released on 24 November 2025, Opus 4.5 delivers stronger reasoning, more capable coding and agentic workflows, and improved long-context performance for enterprise use cases.

What’s new in Claude Opus 4.5

  • Enhanced reasoning and coding across complex tasks, including software development, data analysis and multi-step problem solving, with leading performance on coding and agent benchmarks.
  • Stronger agentic capabilities, enabling agents to work more effectively with spreadsheets, documents and other tools as part of end-to-end workflows.
  • Improved context handling and memory for long-running, multi-step interactions where consistency and reliability are critical.

Benefits for Zeaware Avalon customers

  • Smarter agents and workflows for coding, policy Q&A, analytics and operations, powered by a frontier-grade reasoning model.
  • More reliable performance and flexibility through upgraded backend model support within Avalon’s governed, enterprise-ready framework.
  • Opportunities for more advanced automation using Opus 4.5’s improved planning, tool-use and long-horizon task handling.

Getting started

Claude Opus 4.5 can be added within the service or model settings in your Avalon environment. To explore the new capability, you can:

  • Re-run existing workflow tests and compare results when switching to Opus 4.5.
  • Use more complex, multi-step tasks or agent flows that benefit from stronger reasoning and tool-use.
  • Review workflows that previously required manual oversight, as Opus 4.5 may reduce orchestration complexity.

What’s next

Our roadmap includes exposing additional Claude Opus 4.5-powered features across Avalon, including enhanced code-aware agents, richer analysis workflows and more capable autonomous and semi-autonomous agent interactions. We will share updates as these capabilities become available.

For a demonstration of Claude Opus 4.5 within Avalon or guidance on upgrading your workflows, please contact the Zeaware team.

]]> We are pleased to announce that Zeaware Avalon now supports Anthropic’s latest flagship model, Claude Opus 4.5. Released on 24 November 2025, Opus 4.5 delivers stronger reasoning, more capable coding and agentic workflows, and improved long-context performance for enterprise use cases.

What’s new in Claude Opus 4.5

  • Enhanced reasoning and coding across complex tasks, including software development, data analysis and multi-step problem solving, with leading performance on coding and agent benchmarks.
  • Stronger agentic capabilities, enabling agents to work more effectively with spreadsheets, documents and other tools as part of end-to-end workflows.
  • Improved context handling and memory for long-running, multi-step interactions where consistency and reliability are critical.

Benefits for Zeaware Avalon customers

  • Smarter agents and workflows for coding, policy Q&A, analytics and operations, powered by a frontier-grade reasoning model.
  • More reliable performance and flexibility through upgraded backend model support within Avalon’s governed, enterprise-ready framework.
  • Opportunities for more advanced automation using Opus 4.5’s improved planning, tool-use and long-horizon task handling.

Getting started

Claude Opus 4.5 can be added within the service or model settings in your Avalon environment. To explore the new capability, you can:

  • Re-run existing workflow tests and compare results when switching to Opus 4.5.
  • Use more complex, multi-step tasks or agent flows that benefit from stronger reasoning and tool-use.
  • Review workflows that previously required manual oversight, as Opus 4.5 may reduce orchestration complexity.

What’s next

Our roadmap includes exposing additional Claude Opus 4.5-powered features across Avalon, including enhanced code-aware agents, richer analysis workflows and more capable autonomous and semi-autonomous agent interactions. We will share updates as these capabilities become available.

For a demonstration of Claude Opus 4.5 within Avalon or guidance on upgrading your workflows, please contact the Zeaware team.

]]>> What Enterprises Get Wrong About RAG https://live.cms.zeaware.net/zeaware/blog/what_enterprises_get_wrong_about_rag_performance Fri, 21 Nov 2025 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/what_enterprises_get_wrong_about_rag_performance

RAG Is Not Dead, It’s Evolving

With the rise of advanced function calling, agent frameworks, and multi-step LLM workflows, many organisations are asking:

“Do we still need RAG?”

Some even assume function calls like search_content() make RAG obsolete.

They don’t. In fact, RAG has become more important, because modern AI systems now blend three retrieval modes:

  • Developer-controlled RAG (deterministic, governed evidence)
  • Model-initiated content search (dynamic exploration)
  • Governance-directed hybrid retrieval (the correct blend of both)

The core issue is no longer RAG vs function calling. It is: Who controls knowledge injection into the model: the developer, the model, or the governance layer?

This shift is fundamental for enterprise-grade AI systems.

Many Enterprises Misunderstand RAG

Many organisations still implement RAG as:

“Split documents → embed chunks → keyword + vector search.”

But high-performance RAG is actually a governed evidence pipeline with:

  • structural segmentation
  • metadata extraction
  • entity linking
  • hybrid vector + keyword retrieval
  • effective-dated filters
  • scoring and ranking
  • version control
  • lineage tracking
  • governance gates
  • audit trails

RAG is not a “search mechanism”. It is a content governance mechanism. When RAG underperforms, it’s rarely an embedding issue. It’s because the enterprise designed a search feature, not an evidence pipeline.

Why Function Calling Makes RAG More Important

Function calling introduces powerful new behaviour:

  • calling internal systems
  • invoking workflow logic
  • running calculations
  • executing approvals
  • retrieving structured data
  • performing validation

But with this power comes a risk that the model may attempt to execute logic on the wrong information.

For example:

  • calling a lookup tool before retrieving the correct clause
  • making decisions on outdated or incorrect source material
  • misinterpreting vague queries and calling the wrong function
  • pulling broad or irrelevant search results

RAG provides the governed, correct, versioned evidence that logic must operate on. The more you rely on function calling, the more you need governed retrieval to ensure the model is using the right evidence.

Three Retrieval Modes

Modern enterprise agents require three different retrieval approaches.

1 - Developer-Controlled RAG

Best for:

  • compliance
  • policy interpretation
  • risk decisions
  • claims logic
  • regulated content
  • version-controlled documents
  • anything where correctness is more important than creativity

In this mode, the developer or platform decides:

  • what content is allowed
  • how it’s chunked and structured
  • what can be injected into context
  • how retrieval is filtered
  • which versions are permitted

This can help increase predictability and trust - if done well - or lead to poor quality responses and apparent hallucinations if implemented poorly.

2 - Model-Initiated Search

Best for:

  • exploratory queries
  • broad discovery
  • multi-topic questions
  • cases where user intent is unclear
  • “find anything relevant” tasks

Here the model decides:

  • when it needs more information
  • what to search for
  • how to form the query

This introduces flexibility, but also risk.

Model-controlled search can:

  • misinterpret intent
  • generate poor or overly broad queries
  • retrieve irrelevant or excessive material
  • assume enough knowledge from existing context

Powerful, but not robust on its own.

3 - Governance-Directed Hybrid

This is where Zeaware Avalon is focused.

In this mode:

  • the governance layer influences how retrieval is actually performed
  • the platform can refine, override, or block the model’s choices
  • deterministic filters and metadata rules are enforced
  • evidence is validated before injection
  • context must pass governance gates

This mode ensures:

  • model flexibility
  • platform safety
  • deterministic results
  • resilient, trustworthy answers

This is the pattern enterprise AI will adopt at scale.

How Governance Controls Retrieval (Not the Model)

AI governance is not just logging, security, or risk reporting. At Zeaware, we define a component of AI governance as the systematic control of how knowledge is injected into reasoning and used in outcomes.

The governance layer decides:

  • when strict RAG is required
  • when exploration via search is allowed
  • when the model’s retrieval choice must be overridden
  • which sources are permitted
  • which metadata filters must apply
  • how retrieval aligns with business rules
  • what evidence is admissible
  • what conditions must be met before synthesis

Governance transforms retrieval from a guess into a reliable system.

This is why we say:

  • RAG gives the model facts.
  • Tools give the model abilities.
  • Governance ensures the system produces correct and consistent outcomes.

The Future: Retrieval as a Governed Spectrum

In the next generation of enterprise AI systems, retrieval will not be one technique.

It will be a governed spectrum, enterprises that treat retrieval as a single mechanism will see unpredictable behaviour.

Enterprises that treat retrieval as a governed spectrum will see:

  • safer AI
  • more accurate answers
  • fewer hallucinations
  • better user trust
  • more robust workflows
  • more scalable patterns

This is where the market is moving.

RAG Isn’t Competing With Function Calling, They Solve Different Problems

The winning architecture is not one vs the other. It is retrieval by governance, the system deliberately choosing the right retrieval mode for each task.

This is exactly what Zeaware’s Avalon platform is built for: governed evidence pipelines, controlled tool orchestration, and agents that operate reliably and transparently.

RAG isn’t dead. It’s becoming a controlled gateway for knowledge, rather than a search feature.


The views and technical opinions expressed in this article reflect the perspective of the Zeaware engineering team at the time of writing. They are provided for information and educational purposes only and should not be interpreted as formal product commitments, guarantees, or as a substitute for independent architectural or security assessment. Zeaware’s platform is continually evolving, and capabilities, terminology, and recommended patterns may change over time. For specific implementation guidance or to validate suitability for your environment, please contact Zeaware directly.

]]>

RAG Is Not Dead, It’s Evolving

With the rise of advanced function calling, agent frameworks, and multi-step LLM workflows, many organisations are asking:

“Do we still need RAG?”

Some even assume function calls like search_content() make RAG obsolete.

They don’t. In fact, RAG has become more important, because modern AI systems now blend three retrieval modes:

  • Developer-controlled RAG (deterministic, governed evidence)
  • Model-initiated content search (dynamic exploration)
  • Governance-directed hybrid retrieval (the correct blend of both)

The core issue is no longer RAG vs function calling. It is: Who controls knowledge injection into the model: the developer, the model, or the governance layer?

This shift is fundamental for enterprise-grade AI systems.

Many Enterprises Misunderstand RAG

Many organisations still implement RAG as:

“Split documents → embed chunks → keyword + vector search.”

But high-performance RAG is actually a governed evidence pipeline with:

  • structural segmentation
  • metadata extraction
  • entity linking
  • hybrid vector + keyword retrieval
  • effective-dated filters
  • scoring and ranking
  • version control
  • lineage tracking
  • governance gates
  • audit trails

RAG is not a “search mechanism”. It is a content governance mechanism. When RAG underperforms, it’s rarely an embedding issue. It’s because the enterprise designed a search feature, not an evidence pipeline.

Why Function Calling Makes RAG More Important

Function calling introduces powerful new behaviour:

  • calling internal systems
  • invoking workflow logic
  • running calculations
  • executing approvals
  • retrieving structured data
  • performing validation

But with this power comes a risk that the model may attempt to execute logic on the wrong information.

For example:

  • calling a lookup tool before retrieving the correct clause
  • making decisions on outdated or incorrect source material
  • misinterpreting vague queries and calling the wrong function
  • pulling broad or irrelevant search results

RAG provides the governed, correct, versioned evidence that logic must operate on. The more you rely on function calling, the more you need governed retrieval to ensure the model is using the right evidence.

Three Retrieval Modes

Modern enterprise agents require three different retrieval approaches.

1 - Developer-Controlled RAG

Best for:

  • compliance
  • policy interpretation
  • risk decisions
  • claims logic
  • regulated content
  • version-controlled documents
  • anything where correctness is more important than creativity

In this mode, the developer or platform decides:

  • what content is allowed
  • how it’s chunked and structured
  • what can be injected into context
  • how retrieval is filtered
  • which versions are permitted

This can help increase predictability and trust - if done well - or lead to poor quality responses and apparent hallucinations if implemented poorly.

2 - Model-Initiated Search

Best for:

  • exploratory queries
  • broad discovery
  • multi-topic questions
  • cases where user intent is unclear
  • “find anything relevant” tasks

Here the model decides:

  • when it needs more information
  • what to search for
  • how to form the query

This introduces flexibility, but also risk.

Model-controlled search can:

  • misinterpret intent
  • generate poor or overly broad queries
  • retrieve irrelevant or excessive material
  • assume enough knowledge from existing context

Powerful, but not robust on its own.

3 - Governance-Directed Hybrid

This is where Zeaware Avalon is focused.

In this mode:

  • the governance layer influences how retrieval is actually performed
  • the platform can refine, override, or block the model’s choices
  • deterministic filters and metadata rules are enforced
  • evidence is validated before injection
  • context must pass governance gates

This mode ensures:

  • model flexibility
  • platform safety
  • deterministic results
  • resilient, trustworthy answers

This is the pattern enterprise AI will adopt at scale.

How Governance Controls Retrieval (Not the Model)

AI governance is not just logging, security, or risk reporting. At Zeaware, we define a component of AI governance as the systematic control of how knowledge is injected into reasoning and used in outcomes.

The governance layer decides:

  • when strict RAG is required
  • when exploration via search is allowed
  • when the model’s retrieval choice must be overridden
  • which sources are permitted
  • which metadata filters must apply
  • how retrieval aligns with business rules
  • what evidence is admissible
  • what conditions must be met before synthesis

Governance transforms retrieval from a guess into a reliable system.

This is why we say:

  • RAG gives the model facts.
  • Tools give the model abilities.
  • Governance ensures the system produces correct and consistent outcomes.

The Future: Retrieval as a Governed Spectrum

In the next generation of enterprise AI systems, retrieval will not be one technique.

It will be a governed spectrum, enterprises that treat retrieval as a single mechanism will see unpredictable behaviour.

Enterprises that treat retrieval as a governed spectrum will see:

  • safer AI
  • more accurate answers
  • fewer hallucinations
  • better user trust
  • more robust workflows
  • more scalable patterns

This is where the market is moving.

RAG Isn’t Competing With Function Calling, They Solve Different Problems

The winning architecture is not one vs the other. It is retrieval by governance, the system deliberately choosing the right retrieval mode for each task.

This is exactly what Zeaware’s Avalon platform is built for: governed evidence pipelines, controlled tool orchestration, and agents that operate reliably and transparently.

RAG isn’t dead. It’s becoming a controlled gateway for knowledge, rather than a search feature.


The views and technical opinions expressed in this article reflect the perspective of the Zeaware engineering team at the time of writing. They are provided for information and educational purposes only and should not be interpreted as formal product commitments, guarantees, or as a substitute for independent architectural or security assessment. Zeaware’s platform is continually evolving, and capabilities, terminology, and recommended patterns may change over time. For specific implementation guidance or to validate suitability for your environment, please contact Zeaware directly.

]]>> Gemini 3 now supported in Zeaware Avalon https://live.cms.zeaware.net/zeaware/blog/avalon_gemini_3_support Thu, 20 Nov 2025 00:00:00 GMT https://live.cms.zeaware.net/zeaware/blog/avalon_gemini_3_support We are pleased to announce that Zeaware Avalon now supports Google’s Gemini 3 model. This upgrade brings stronger reasoning, multimodal understanding and more capable agentic workflows directly into the platform.

What’s new in Gemini 3

  • Enhanced reasoning and multimodal processing across text, images, video and audio, allowing Avalon to interpret more complex inputs with higher accuracy.
  • Agentic task capabilities with longer context windows and improved tool-use, supporting richer multi-step workflows.
  • Better productivity features for content generation, analysis and design tasks, enabling faster time-to-insight for enterprise users.

Benefits for Avalon customers

  • Smarter summaries, insights and workflows when working with long documents, structured data or multimodal content.
  • More reliable performance and flexibility with upgraded backend model support.
  • Opportunities for more advanced automation through improved reasoning, planning and context handling.

Getting started

Gemini 3 can be added within service settings in your Avalon environment. To explore the new capability, you can:

  • Re-run existing workflow tests and compare results with the new model.
  • Include multimodal inputs such as images and text together.
  • Review agent workflows that previously required manual orchestration, as Gemini 3 may reduce complexity.

What’s next

Our roadmap includes exposing additional Gemini 3-powered features across Avalon, including enhanced analytics, media-rich summarisation and more capable agent interactions. We will share updates as these features become available.

For a demonstration of Gemini 3 within Avalon or guidance on upgrading your workflows, please contact the Zeaware team.

]]> We are pleased to announce that Zeaware Avalon now supports Google’s Gemini 3 model. This upgrade brings stronger reasoning, multimodal understanding and more capable agentic workflows directly into the platform.

What’s new in Gemini 3

  • Enhanced reasoning and multimodal processing across text, images, video and audio, allowing Avalon to interpret more complex inputs with higher accuracy.
  • Agentic task capabilities with longer context windows and improved tool-use, supporting richer multi-step workflows.
  • Better productivity features for content generation, analysis and design tasks, enabling faster time-to-insight for enterprise users.

Benefits for Avalon customers

  • Smarter summaries, insights and workflows when working with long documents, structured data or multimodal content.
  • More reliable performance and flexibility with upgraded backend model support.
  • Opportunities for more advanced automation through improved reasoning, planning and context handling.

Getting started

Gemini 3 can be added within service settings in your Avalon environment. To explore the new capability, you can:

  • Re-run existing workflow tests and compare results with the new model.
  • Include multimodal inputs such as images and text together.
  • Review agent workflows that previously required manual orchestration, as Gemini 3 may reduce complexity.

What’s next

Our roadmap includes exposing additional Gemini 3-powered features across Avalon, including enhanced analytics, media-rich summarisation and more capable agent interactions. We will share updates as these features become available.

For a demonstration of Gemini 3 within Avalon or guidance on upgrading your workflows, please contact the Zeaware team.

]]>>