Copilot Studio adds app building and more controls for enterprise agents

Microsoft is expanding Copilot Studio beyond conversational agents, adding ways to build apps and deterministic workflows alongside them. Its September update, published October 7, also describes controls for tool calls, enterprise knowledge access and agent evaluation.

Several of the additions have different release statuses. Apps and Copilot Managed Runtime are in public preview. Hooks are in preview for agents powered by the GitHub Copilot harness. Foundry IQ integration and the Review panel are generally available. Teams should keep those distinctions visible when planning adoption.

App building alongside agents and workflows

Makers can describe an app in natural language, including its business outcome, intended users, data and actions. Copilot Studio generates a draft app that they can preview and refine in conversation. Microsoft says makers retain access to the underlying code for deeper customization.

The company uses onboarding as an example of combining the different pieces. An app can provide a structured interface, workflows can perform repeatable steps, and an agent can answer questions or handle exceptions. The division gives teams a way to decide which parts of a process need consistent execution and which need interpretation.

Copilot Managed Runtime provides Microsoft-operated hosting, identity, governed data access, lifecycle management, Git-backed versioning and centralized visibility for supported apps. Microsoft describes it as available in public preview across apps created through Copilot Studio, Copilot Cowork and Copilot Code.

Those platform features are intended to support operational control. They do not remove the need to test the generated application against the organization's permissions, data requirements and expected process.

Hooks can intercept a tool call

Hooks bring deterministic logic to specified points in an agent's lifecycle. In the GitHub Copilot harness, Microsoft lists session starts, tool calls and errors among the relevant events.

Makers can use a hook to load required context at session start, inspect or block a tool call before execution, transform a tool's results or respond to failures. Hooks can also collect consistent logs and telemetry.

This is a useful distinction from asking an agent to remember a rule in a prompt. A hook determines when defined logic runs. Microsoft presents that as a way to enforce actions that need to happen each time while leaving the agent room to reason around those controls. Teams still need to decide which rules belong in that logic and verify what happens when a hook rejects an action.

Reusable knowledge and broader evaluation

The Foundry IQ integration lets multiple Copilot Studio agents reuse a knowledge base. Each agent can have knowledge sources and a retrieval scope suited to its purpose. Microsoft says agents can provide grounded responses with citations and can connect through private networking.

Evaluation is also expanding to individual AI nodes and broader automation. The announced checks cover task completion, tool accuracy, configurable safety thresholds and latency. A reusable Custom Graders Library supports criteria across agents and test sets. Generated graders remain specific to the tenant.

The new Evaluation Viewer role lets reviewers see evaluation results without broader maker permissions. That is a practical addition for teams where the person checking an agent's behavior should not have the same access as the person building it.

Readiness checks before publishing

The Review panel surfaces blockers and warnings during building, with explanations and possible actions. Microsoft says evaluation warnings can flag an agent that has never been evaluated or a model change that may make earlier evaluation results less relevant.

The panel can also surface organization-specific policy restrictions. Teams should treat these signals as inputs to a release decision rather than proof that a passing agent will handle every situation correctly.

Microsoft also says its centralized plugin registry contains more than 100 plugins, with Copilot Studio support rolling out over the coming weeks. The broader direction is a shared place for building business interfaces and agent-driven processes. The strongest operational additions in this update are the ones that make behavior inspectable: tool-call controls, repeatable evaluation and visible publishing blockers.