Hi,
The behavior you described is something that can happen when an agent is published to the Microsoft Teams channel, especially when the agent is using external tools such as an MCP server.
When you use “Test your agent” in Copilot Studio, the request is sent directly to the Copilot runtime, so the response time is usually faster.
However, when the agent is used in Microsoft Teams, the request goes through additional services before reaching the model and the tool:
Teams' client → Teams bot service → Copilot runtime → model/tool → response back through the same path.
Because of this extra processing layer, responses in Teams can sometimes be slower than in the Copilot Studio test environment.
Another factor in your case is the MCP server tool. When the model decides to use the tool, the runtime must call the MCP server, wait for the result, and then generate the final answer. This additional round trip can significantly increase response time compared to a simple model response.
The occasional SystemError you see can also occur when a request takes longer than expected or if there is a temporary service retry in the channel integration.
A few things you might consider checking:
-
Whether the MCP tool is being called on every turn
-
If the tool response time is high
-
Testing with and without the tool enabled to compare latency
-
Monitoring logs or tool execution times to see where the delay occurs
Since the same MCP server works faster in other clients, the additional latency is likely related to the Teams channel integration and the tool execution path, rather than the model itself.
Hope this helps.
