Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Agents powered by the GitHub Copilot harness in Microsoft Copilot Studio run each user message as a turn. A turn can include many model calls and tool calls. To keep turns reliable, the service applies limits to how long a turn runs, how many tool calls it makes, and how much conversation history the agent keeps. This article describes those limits and what you see when a limit is reached.
Note
The harness used to build this agent uses usage-based billing. Using, building, testing, and evaluating agents might consume Copilot Credits. Learn more in Overview of usage-based billing.
Turn and tool time limits
The values in the following table are the service defaults. Microsoft manages these values and can change them.
| Limit | Default value | What happens when reached |
|---|---|---|
| Turn duration | 90 minutes | About 60 minutes into a turn, the agent is steered to wrap up and give a final answer. At 90 minutes, the service stops the turn. |
| Tool calls per turn | 200 | The agent stops calling tools and the turn ends. |
| Tool call timeout | 120 seconds | The service cancels the tool call and returns an error result to the agent. |
| Workflow tool timeout | 660 seconds (11 minutes) | The workflow tool call times out and the agent receives an error. |
| Connected agent timeout | 5,400 seconds (90 minutes) | The call to the connected agent times out and the agent receives an error. |
| External agent timeout | 300 seconds (5 minutes) | The call to the external agent times out and the agent receives an error. |
| Messages kept in the conversation window | 100 messages | The agent drops the oldest messages from its working history, so it can forget details from early in a long conversation. |
The tool call timeout is the default for an ordinary tool call, such as a connector call or a knowledge search. Workflow tools, connected agents, and external agents use their own longer timeouts.
If a request regularly reaches one of these limits, break the work into smaller requests. To look up a reference code that appears with a failed turn, such as MCS-4063, see Error codes reference for agents.
Streaming limits in Microsoft Teams and Microsoft 365 Copilot
Agents stream responses as the agent writes them. In Microsoft Teams, streaming has the following limits:
- Personal chats only: Streaming applies to one-on-one personal chats. In group chats and channels, the agent sends answers as ordinary messages.
- Update rate: Teams accepts about one streaming update per second. The service groups text together so each update stays within that limit.
- Two-minute limit: Teams ends a streamed response after two minutes, even if the agent is still working.
- Stop: If a user selects Stop on a streaming response, Teams ends the stream.
If a response runs past the two-minute limit, the service sends the complete answer as a regular message instead of finishing the stream. The text that already streamed can remain on screen, so the answer can appear twice. If a long tool call uses up the time before any answer text arrives, you might see a stopped response followed by the answer in a separate message. Either way, the full final message is still delivered.
Microsoft 365 Copilot doesn't use the Teams streaming session, so the Teams limits in this section don't apply to it.