Microsoft Foundry Update: Voice Agents in Preview; Routines, A2A and Tool Search Now GA

On September 24, 2026, Microsoft published a batch of updates to Microsoft Foundry, its platform for building and running AI agents on Azure. The announcement on the Azure blog covers three areas: more models to choose from, agents that can speak and listen, and tools for measuring and improving an agent after it goes live.

For a business that already runs on Microsoft 365 and Azure, Foundry is often the default place to build an agent, whether in-house or through a software development partner. The difficulty with announcements like this one is that features at very different stages of maturity are presented side by side.

This article separates what Microsoft says is generally available from what is in public preview and what is still to come, then looks at what the changes mean for a company planning or running an agent.

What Microsoft says is available now

According to the post, the following are available or generally available as of September 24, 2026:

• Models: the GPT-6 family from OpenAI, including GPT-6 Sol and GPT-6 Luna, and Claude Opus 5.5 from Anthropic are described as now available in Microsoft Foundry.
• Tool search in Toolboxes, which lets an agent find the right tool without loading every tool definition up front.
• Agent-to-Agent (A2A), which lets agents call other agents, including other Foundry agents, through an open, standardized protocol.
• Routines in Foundry Agent Service, which run agents on a schedule, after a delay, or in response to an event.

Microsoft also notes that Toolboxes, its way of giving agents governed access to tools, were already generally available for hosted agents.

On tool search, the post reports Microsoft's own measurement: input-token consumption fell by more than 60 percent for a toolbox of 50 tools and by over 97 percent for a toolbox of 1,000 tools. These are vendor figures for a specific setup, not a promise about your bill, but the direction matters. Every tool definition an agent carries is text the model has to read on each request, and you pay for that reading.

What is in public preview

Several of the more eye-catching items are previews, which means they can change and are generally not what you want underneath a customer-facing launch. The post lists these as in public preview:

• Voice agents in Foundry Agent Service.
• Long-running resilience, which lets the work of a hosted agent continue after a request disconnects.
• Toolboxes for prompt agents, extending the feature beyond hosted agents.
• Insights in Foundry, which analyzes production traces to surface recurring issues.

Voice agents in more detail

Microsoft says voice agents are built, deployed, observed and evaluated through the same platform as text agents, and that developers can reuse the tools, knowledge and governance already configured for their other Foundry agents. The post states support for more than 80 languages and more than 140 locales, and lists deployment targets as web channels, Microsoft Teams, Teams Phone, and inbound and outbound telephony through Twilio. The announcement gives no pricing for voice.

What is announced but not yet confirmed as shipped

Three evaluation and tuning features were described on September 24 as becoming generally available "later this month": a rubric evaluator for defining what a good answer looks like, generation of evaluation datasets from synthetic data and from production traces, and an agent optimizer. Microsoft says the optimizer tests changes to instructions, tools, skills and model choice and shows which changes improve results before anything ships.

We have not confirmed from a Microsoft source that these three reached general availability by the end of September, so treat them as announced until you see them in your own tenant or in Microsoft's documentation. Separately, the post says preview support for a new AI Gateway tier in Azure API Management is coming in October 2026. That item is both future-dated and a preview.

Security and governance additions

The post also describes network egress controls, which govern the outbound destinations a hosted agent can reach, with an audit mode that lets a team see what a policy would block before enforcing it. The announcement does not label this feature as preview or generally available, so check its status in the documentation before relying on it.

The idea is worth understanding regardless of status. An agent that can call tools and browse internal systems can also send data somewhere it should not, either through a mistake or because a malicious document told it to. Limiting where an agent can connect is one of the few controls that still works when the model itself has been misled.

What this means for your business

Three practical points follow from the announcement. The first is that model choice inside one platform is becoming normal. With OpenAI and Anthropic models offered side by side in Foundry, a business can switch the model behind an agent without changing where its data, identity and billing live. That reduces lock-in to a single model maker, though it increases dependence on the platform.

The second is that scheduled and event-driven agents are now a supported feature rather than something a developer assembles by hand. Routines make it realistic to run an agent every morning against new orders or whenever a support ticket arrives. That is a different thing from a chatbot waiting for a question, and it needs different safeguards. Our guide to AI agents, chatbots and workflow automation explains the distinction.

The third is cost. Microsoft's emphasis on tool search, evaluation and optimization is an acknowledgment that agents can be expensive to run and hard to judge. If a vendor quotes you a build price for an agent without a plan for measuring quality and token spend after launch, the quote is incomplete.

What to do next

If you already run agents in Foundry, ask your team two questions. How many tools does each agent load, and would tool search reduce that? And do any workflows currently rely on custom scheduling code that Routines could replace? Both are generally available according to Microsoft, so they are reasonable candidates for production use after testing.

If you are planning a voice agent, a public preview is a good place to prototype and a risky place to commit. Build a pilot, test it with real accents, background noise and interruptions, and keep a human handoff path. Hold back launch dates and contractual promises until Microsoft announces general availability and publishes pricing you can model.

If you are not on Azure at all, there is no reason to move because of this announcement alone. Comparable building blocks exist on other clouds and directly from model providers. The better starting point is the process you want to automate, the systems it touches and the level of human approval it needs; the platform choice follows from those.

Conclusion

Microsoft's September 24, 2026 Foundry update makes tool search, agent-to-agent calls and scheduled Routines generally available, adds GPT-6 and Claude Opus 5.5 models, and puts voice agents, long-running resilience and Insights into public preview. Evaluation and optimization features were announced for later in September, and an AI Gateway preview is due in October 2026.

The useful habit is to read each item's status before planning around it. If you would like a second opinion on whether an agent fits a process in your business, and on which platform, you can request a quote and describe what you want to automate.