Microsoft gives Copilot Studio apps, tighter agent controls and new evaluations

Microsoft’s September Copilot Studio update brings apps, workflows and agents into one building environment, while adding new ways to control and evaluate agents. For teams putting AI into business processes, the notable shift is from assembling the parts to managing more of the build and readiness checks in one place.

Microsoft AI Watch analysis

What happened

Microsoft says apps are now in public preview in Copilot Studio. Makers can describe an app in natural language, refine the generated draft conversationally and still reach its underlying code for deeper changes. Microsoft also says Copilot Managed Runtime, its hosting and governance environment for these apps, is in public preview.

The update adds agent hooks in preview for the GitHub Copilot harness, connecting predictable actions to events such as session starts, tool calls and errors. Foundry IQ integration is generally available, and evaluations now cover automation as well as agent conversations. Microsoft’s Copilot Studio update sets out the full changes.

Our top picks

• Build apps from a prompt
  Makers can generate and refine a draft app in natural language, with access to the code for further customisation.
• Put predictable checks around agent actions
  Hooks can inspect, modify or block tool calls, and respond to starts, results and errors.
• Reuse enterprise knowledge
  Foundry IQ knowledge bases can serve multiple agents, with citations intended to help users check answers.
• Evaluate more than the conversation
  Teams can assess individual AI nodes and broader automations, including task completion, tool accuracy, safety and latency.
• Share evaluation results without maker access
  A new Evaluation Viewer role lets reviewers see results without broader permissions.
• Spot readiness issues earlier
  The generally available Review panel surfaces blockers, warnings and relevant policy restrictions before publication.

Why it matters

Business processes often need both flexibility and consistency. An agent can adapt to an unusual request, but a tool call may still need a reliable check every time. Hooks offer one way to put that boundary in place, while expanded evaluations give teams more specific measures than a general impression that the agent seemed helpful.

This is a substantial set of building and governance changes, but it is not proof that the resulting apps or agents will perform well in production. Public previews remain previews, and the practical test is whether teams can use these controls to catch problems before work reaches customers or staff.

Our read

The most useful change is the combination: build the interface, define repeatable work and add agents where judgement is needed, then evaluate the result. That is a more convincing pitch than asking businesses to trust an agent because it has a polished demo. Teams should look closely at the new checks, especially whether they reflect the consequences of the task rather than simply producing reassuring scores.

What to watch

• Which Copilot Studio app and runtime features move beyond public preview.
• Whether hooks and evaluations catch meaningful failures in real workflows.
• How broadly Foundry IQ integration and the plugin registry become available.

Discussion spark: For business AI agents, which should be the priority: strict controls on what they can do, or broader evaluations of whether they achieve the right outcome?

not affiliated with or endorsed by Microsoft