Ship agents faster with expanded model choice, voice agents, and continuous optimization

Your business model isn’t a one-time decision; it needs to evolve continually. Embracing a new model should propel your business forward rather than force your team to rebuild everything from scratch. As new models emerge, organisations should have the flexibility to choose what’s best for each unique workload while relying on a foundational backbone that can adapt as their needs change.

Enter Microsoft Foundry—a versatile and model-agnostic framework designed for developing and managing agents. With Foundry, teams can adopt the latest models as they roll out, all while safeguarding their existing investments in enterprise systems, know-how, tools, and protocols that enhance agent efficiency.

However, implementing a new model is just the beginning. Foundry empowers teams to continuously enhance their agents. By harnessing production data, they can assess and fine-tune instructions, skills, tools, and model selections based on performance metrics like quality, latency, and cost. This iterative process, known as the hill-climbing approach, encourages teams to observe, assess, optimise, validate, and repeat, all while keeping human oversight at the forefront.

Today, we’re making further advancements to ensure teams can build and consistently refine agents that offer tangible business value:

• Expand your model selection as cutting-edge AI technology develops.
• Create voice interactions and manage extended tasks using the same foundational agent technology.
• Provide agents with the knowledge and resources they need to utilise the right capabilities effectively.
• Transform production insights into ongoing improvements that enhance agent quality, latency, and cost efficiency.

These improvements enable teams to focus less on constructing infrastructure and allocate more time to innovation.

A prime example of these capabilities in action is Fashable. They leverage Foundry’s models, agents, knowledge, and tools to assist fashion brands in translating trends and concepts into clothing designs, virtual try-ons, and promotional visuals. By integrating these workflows with existing brand systems, Fashable has slashed the time needed for new product development from months to mere weeks, with one client experiencing a remarkable 60% reduction in their budget for physical samples.

Selecting and assessing frontier AI models in Microsoft Foundry

The field of frontier AI is advancing at breakneck speed. Breakthrough features are emerging constantly—often in just days. This week, we see that momentum accelerating with the full launch of the GPT-6 family from OpenAI, which includes GPT-6 Sol and GPT-6 Luna, along with Claude Opus 5.5 by Anthropic. The significance of this moment isn’t pinpointed on a single model, but rather on the swift pace and extensive range of innovation taking place across various fronts simultaneously.

For developers and businesses alike, the selection of a model is no longer a one-off decision at the project’s outset. Instead, it’s a continuous advantage. The ideal model for a coding agent might differ from what works best for research, customer support, or high-volume inference tasks. Moreover, the “best” solution will change continuously as new models appear. That’s where Microsoft Foundry truly excels.

Foundry consolidates models from top AI providers into an open platform. This enables teams to evaluate them with their own data and workloads, weigh the trade-offs in terms of quality, latency, and costs, and deploy the right models without needing to redesign their AI strategy around a single provider. With the introduction of the GPT-6 family and Claude Opus 5.5, choices expand further, allowing teams to pick from a wider array of models tailored to their specific needs.

Develop and manage voice agents using Foundry Agent Service

Voice is swiftly becoming a primary mode of interaction for AI agents. However, many voice experiences still rely on layering speech over text-based agents, necessitating separate components for speech processing, orchestration, monitoring, and deployment. With voice agents in Foundry Agent Service, vocal interaction is natively integrated into the agent foundation—built, deployed, monitored, assessed, and refined using the same platform, API, and SDK as all other Foundry agents.

Currently in public preview, voice agents represent a new type that supports both prompt-based and hosted agents. Developers can select from leading voice models—such as GPT Realtime, Azure Realtime, MAI, and others—to optimise for latency, reasoning depth, and cost for each context. Alternatively, they can bring their own models. These tools and deployment pathways are created specifically for verbal interaction, ensuring expressive voices, seamless turn-taking, and smooth adjustments if interruptions occur, covering over 80 languages and 140 locales. For teams wanting to establish a unique identity, custom vocabularies and voices can be fine-tuned, alongside the option of incorporating a photo or video avatar to create a distinctive brand presence. Developers can leverage the existing tools, insights, and governance set up for other Foundry agents, with deployment channels including web platforms, Microsoft Teams, Teams Phone, and Twilio’s inbound and outbound telephony services.

The workflow for developers remains streamlined across the tools already in use:

• AZD AI offers a consistent, source-controlled process for configuring and deploying voice agents.
• Foundry Toolkit in Visual Studio Code (coming soon) will empower developers to create, run, interact with, and debug voice agents all within their development environment.
• Voice-agent observability brings tracing, monitoring, transcripts, and voice-specific evaluation to spoken interactions.

At Fashable, our goal is to make advanced AI accessible to brands and manufacturers of all sizes, no matter their technical expertise. As we unveil new agent capabilities, voice plays a crucial role in that journey, fostering interactions that feel more natural, conversational, and accessible. People don’t think in prompts; they think in conversations. Utilising voice agents in Foundry with MAI-Voice helps bridge that gap.

—Orlando Ribas Fernandes, CEO, Fashable

Create robust hosted agents for extended tasks

In a real business setting, tasks often don’t complete within a single request. A research task might involve multiple sources, a report could require several tools and processing phases, and an agent may need to pause for external systems or human approval. The long-running resilience in Foundry Agent Service, currently in public preview, allows the work of hosted agents to continue seamlessly after a request disconnects, with the ability to recover if the hosting process unexpectedly halts.

The Microsoft Agent Framework enhances these features with robust workflow states and checkpointing. If a resilient background workflow is interrupted, the Agent Framework can resume from its latest durable checkpoint, while Foundry ensures that the hosted response remains intact.

Additionally, the Microsoft Agent Framework incorporates the Agent Channel for messaging integrations, AG-UI support for interactive applications, CodeAct and Hyperlight containers for isolated code execution, and episodic procedural memory to reuse successful task patterns. Learn more about the latest advancements in the Microsoft Agent Framework.

To create and deploy these long-running agents, developers can utilise the Foundry Toolkit for Visual Studio Code, which supports hosted agents as well as the Microsoft Agent Framework. The new Foundry dev pack simplifies the onboarding process by consolidating the Foundry developer toolchain into a single installer.

Provide agents with only the resources they need when they need them

Agents in an enterprise environment don’t need to access all documents, tools, and procedures for every interaction. Overloading them with everything from the start adds latency and costs while making their behaviour less predictable. Instead, agents should be able to access the necessary resources on demand as the task progresses.

• Toolboxes allow agents to access tools and reusable procedures on demand, with opt-in discovery and tool-level safety mechanisms. This means teams can enable certain functionalities without requiring them to be loaded during every single call. Toolboxes are already available for hosted agents and are now expanding to prompt agents in public preview.
• Tool search within Toolboxes, now widely available, helps agents locate the right tools without needing to load the entire catalogue upfront. In an internal evaluation that involved more than 44,000 tools and 7,000 queries, tool search cut input-token consumption by over 60% for a toolbox of 50 tools and by over 97% for a toolbox of 1,000 tools. These results were compared with a prompt-cached baseline that loads the full tool catalogue initially.
• Agent-to-Agent (A2A), now generally available, allows agents to communicate with each other, including other Foundry agents, using a standard protocol. This process enables teams to connect agents without constructing custom point-to-point integrations, while maintaining security and governance through centralized credentials and policy enforcement as their systems expand.
• Routines in Foundry Agent Service, also available now, enable teams to run automated agents according to a schedule, after a set delay, or triggered by specific events. There’s no longer a need for teams to create their own schedulers, queues, and identity frameworks just to initiate an agent at the right time.

Collectively, these features assist teams in minimising unnecessary context and custom infrastructure, while offering greater control over what agents can access and when.

Transform production evidence into ongoing improvements

Launching an agent is just the first step. Once real users start utilising it, teams must discern what factors impact quality, latency, and costs, as well as identify which changes can enhance performance without causing regressions. This has often involved a patchwork of traces, dashboards, and on-the-fly tests. Foundry streamlines these processes into a continuous improvement framework, enabling teams to turn production data into actionable insights: observe, understand, evaluate, optimise, validate, and repeat.

• Insights in Foundry, which is currently in public preview, helps teams understand production dynamics. It analyses production traces to highlight recurring issues and those teams might not have previously identified. For each issue flagged, it compiles supporting evidence, suggests probable causes, and offers actionable recommendations, guiding teams on where to concentrate their efforts before a minor issue escalates into a major one.

• Rubric evaluator, which will be available later this month, aids teams in defining successful outcomes based on the issues identified. It converts an agent’s prompts, traces, and requirements into clear success measures that can be systematically tracked.
• Synthetic and traces to dataset generation, also set to be available this month, creates evaluation data that aligns with those success metrics. This correlation is based on actual production traffic, ensuring that testing matches genuine customer scenarios, including those highlighted by Foundry insights.
• Agent optimizer in Foundry Agent Service, will also be launched this month. It utilises those assessments, datasets, and constraints effectively, testing modifications to instructions, tools, skills, and model options, revealing which adjustments truly enhance results before deployment.

Together, these features facilitate the transition from production to production, allowing improved agents to deploy back into real-world usage, where Foundry insights continuously monitor their performance for further enhancements. Each iteration refines the agents, ensuring they inch closer to excellence with time.

The Agent optimizer feature in Foundry Agent Service transforms evaluation insights into operational datasets and optimisations, thereby simplifying the process of identifying and addressing performance challenges. This systematic feedback loop aligns closely with NTT DATA’s AgentOps with Harness approach—a fundamental aspect of our Smart AI Agent® concept—simultaneously advancing agents from proof of concept to trusted, scalable production.

—Takashi Okamoto, AI Technology Strategist, Global AI Office, NTT DATA Group Corporation

Regulate AI agents through identity, lifecycle, and network oversight

As agents take on increasing responsibilities, organisations must manage them as corporate assets, assigning identities, owners, policies, lifecycle controls, and tracking their activities.

Foundry agents are now equipped to honour governance operations initiated in Microsoft Entra and Agent 365. Actions such as blocking, disabling, deleting, restoring, or reallocating ownership are enforced directly by the Foundry runtime rather than merely recorded in a directory.

Network egress controls regulate the outbound destinations accessible to a hosted agent. Teams can define rules for allowance and denial, adjust outbound headers, or redirect requests prior to their exit from the Foundry-managed sandbox. An audit mode permits teams to assess policies before activation, logging every decision in Azure Application Insights.

In October, Microsoft Foundry will expand its integration with Azure API Management, previewing support for a new AI Gateway tier. This enhancement will assist organisations in incorporating central governance into the Foundry developer experience. This means organisations can adopt a hub-and-spoke strategy, where platform teams govern model access through AI Gateway, making approved models available to developers via Admin Connected Models. Developers can then utilise these governed models in familiar Foundry workflows for discovering, experimenting with, developing agents, and evaluating them. Additionally, developers will have the option to connect with models across other resources or external sources directly from the playground. Existing integration with traditional Azure API Management tiers will continue to be supported.

Managing agents as corporate assets also entails confirming that they comply with safety and policy standards. The new open-source run-assert-eval Skill incorporates this verification process into a singular workflow.

It sequentially runs three of Microsoft’s open-source tools:

1. Clarity identifies potential risks developers may overlook.

2. ASSERT converts requirements into measurable assessments.

3. Agent Control Specification (ACS) applies focused runtime controls when an agent does not meet standards.

Once completed, the Skill re-evaluates to confirm that the agent has improved without limiting valid actions, providing teams with evidence that all policies are functioning as intended.

Begin your journey with Microsoft Foundry

The latest features in Microsoft Foundry offer developers a flexible path from a simple prompt to a fully operational production agent. There’s no need for teams to piece together disjointed development tools, runtimes, gateways, schedulers, and governance measures.