Microsoft Foundry Puts AI Agents on a Schedule and Under Watch

Microsoft published its September roundup of updates to Microsoft Foundry, its platform for building and running AI agents. The list is long. But the theme for DevOps teams is clear: Agents are being treated less like chat experiments and more like production workloads that need schedules, cost controls, and observability.

Three capabilities stand out. Routines and Tool Search are now generally available. And Insights, which analyzes agent traces to find recurring problems, is in public preview.

Agents That Run Without Being Asked

Routines let teams start an agent from a trigger. That can be a timer, a recurring schedule, a GitHub issue event, or a supported Microsoft Teams channel event. The configuration, identity, and run history stay with the Foundry project.

In practice, that removes a chunk of plumbing. Teams no longer have to build their own schedulers, queues, and identity handling just to kick off an agent at the right time.

Microsoft is clear about the limits. A routine starts one agent. It doesn’t replace multi-agent orchestration or approval workflows. And a routine that starts a job doesn’t guarantee the job finished or that it delivered its output. Nick Brady, principal product manager for developer experience at Microsoft, recommends starting with a manual run that sends nothing, reviewing the sources, and only then turning on the schedule.

Mitch Ashley, vice president and practice lead for CIO & Technology Buyers and Software Lifecycle Engineering at The Futurum Group, said that caution matters. “A scheduled agent runs when nobody is watching,” Ashley said. “A routine that starts and never delivers fails silently, so teams need completion and delivery checks in place before they turn on a recurring schedule.”

Loading Tools Only When They’re Needed

Tool Search addresses a problem many teams have hit. Every tool definition an agent can use usually gets loaded into each model request. Connect enough tools and those definitions eat into the context window before the agent does any work. That drives up input-token costs and leaves less room for the actual task.

With Tool Search, a Foundry Toolbox exposes two operations, tool_search and call_tool. The agent discovers the definitions it needs as the work unfolds.

Microsoft says an internal evaluation, using a public benchmark of more than 44,000 tools and 7,000 queries, cut input-token use by more than 60% for a 50-tool toolbox and by more than 97% for a 1,000-tool toolbox. Those are Microsoft’s own numbers, measured against a prompt-cached baseline. But independent tests point the same way. In a synthetic test by integration platform Nango, discovery with two pinned tools used 69% fewer input tokens per task than loading all 57 tool definitions upfront.

One caution: discovering a tool isn’t permission to use it. In Microsoft’s example, application code restricts the agent to approved repositories, paths, and read-only methods. Teams still need to enforce permissions where tools execute.

Agent-to-Agent (A2A) support is also generally available. It lets Foundry agents call other agents through an open protocol instead of custom point-to-point integrations.

Finding Problems in the Traces

Insights is the piece most likely to interest operations teams. It analyzes production traces, groups recurring behavior into findings, and attaches supporting evidence, a likely cause, and a proposed next step.

In Brady’s own test workflow, Insights flagged oversized prompts being resubmitted after rate-limit errors, sample-heavy prompts that wasted context, and an unconfirmed delivery receipt. None of those would crash an agent. All of them cost money or erode reliability over time.

Microsoft doesn’t apply fixes automatically. The recommended approach is to inspect one finding, check its linked trace, and test the proposed change against the same input before shipping it.

Microsoft also targeted general availability of a rubric evaluator, trace-to-dataset generation, and an Agent Optimizer for later in the month. Together, they’re meant to form a loop: observe, evaluate, optimize, validate and repeat. That loop will look familiar to anyone who has run a CI/CD pipeline.

Ashley sees observability as the real constraint. “What an organization can observe and prove sets the pace of agent deployment more than model capability does,” he said. “Trace analysis that names recurring problems and likely causes narrows that gap. The next test is whether teams route those findings through the same release gates they use for code.”

Models, Cost and Resilience

On the model side, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are generally available, with a wide spread in pricing. Input tokens run $10 per million for Astra, $2 for Sol, and $0.10 for Luna on Global Standard short-context requests. That gap makes model selection a cost decision, not just a quality one. Anthropic’s Claude Opus 5.5 is also available in Foundry.

Model Router added preview session affinity, which asks the router to prefer the same model across related requests. It also returns metadata showing which model handled each attempt and when fallback occurred. That makes routing decisions easier to audit.

Two previews matter for production. Voice agents arrive as a new agent type. And long-running resilience lets hosted agents keep working after a client disconnects and recover if the hosting process stops. Microsoft notes that application-level checkpointing and idempotency still matter. Reconnecting doesn’t make repeated external actions safe.

Check the SDK and CLI changes before upgrading. The azure.ai.agents extension for the Azure Developer CLI now requires runtime definitions in azure.yaml and rejects legacy discovery. The hosted-agent runtime also defaults to 100% trace sampling in Azure Monitor, which can add up fast.

Why It Matters

Most of these updates aren’t about making agents smarter. They’re about making them easier to schedule, cheaper to run, and easier to troubleshoot. That’s what it takes to move an agent from a demo to something an operations team will support.

And the responsibility doesn’t shift. The teams that build agents still own what those agents do. Foundry now gives them better tools to see it.