Multi-agent setup: answers generated correctly but Microsoft 365 Copilot returns "Sorry, I wasn't able to respond to that" after the first 1–2 turns

uehara 0 Reputation points
2026-09-09T14:16:45.7766667+00:00

Hi everyone,

 

I'm running into a frustrating issue with a multi-agent setup and I'd really appreciate any insight — even partial information or things to rule out would help a lot.

 

My setup

 

Multi-agent architecture in Copilot Studio Agent A (parent): acts purely as a router. Its instructions explicitly tell it not to reason or search on its own, and it has no knowledge sources attached at all. Agent B (child/connected): holds the actual knowledge. It uses SharePoint documents uploaded as knowledge (Dataverse-backed upload, not link-based). A calls B, and B performs the knowledge search and generates the answer.

 

What works

 

In the Copilot Studio test pane, everything works perfectly. B searches the knowledge and returns a complete, accurate answer every single time.

 

What doesn't work

 

When I use the same agent in Microsoft 365 Copilot, the first one or two turns return the answer correctly. After that, every subsequent turn returns:

 

"Sorry, I wasn't able to respond to that. Is there anything else I can help you with?"

 

The confusing part

 

When I check the Activity map / Activity tab in Copilot Studio for those failed turns, the answer IS being generated correctly. The knowledge search runs, and a full, well-formed response exists in the activity log. It just never reaches the user in Microsoft 365 Copilot — the generic error message is shown instead.

 

So this doesn't look like a generation failure or a knowledge retrieval failure to me. It seems like something between the generated response and what Microsoft 365 Copilot actually renders.

 

What I'd like to know

 

Has anyone seen this specific pattern, where the response is confirmed generated in Activity but Microsoft 365 Copilot still shows the generic error? Is this a known limitation with connected/child agents in the Microsoft 365 Copilot channel specifically? Could response length or formatting (my answers contain markdown tables and multiple sections) cause the M365 renderer to drop the response? Could conversation state size or throttling explain why it only fails after the first couple of turns? Is there a better place than Application Insights to see why the M365 channel discarded a response that was successfully generated?

 

Any information at all would be

Microsoft Copilot | Microsoft 365 Copilot | Development
0 comments

1 answer

Sort by: Most helpful
Most helpful Newest Oldest
  1. Sayali-MSFT 6,476 Reputation points Microsoft External Staff Moderator
    2026-09-10T09:06:23.8533333+00:00

    Hello uehara,
    Since the complete answer is visible in the Copilot Studio Activity map, the child agent and knowledge search appear to be completing successfully. The failure is likely occurring after orchestration, when the response is returned to or processed by the Microsoft 365 Copilot channel.

    A few things worth testing:

    • Start a new conversation whenever the failure begins. If this restores functionality for another one or two turns, check for a ConversationStateTooLarge condition or accumulated multi-agent context.
    • Temporarily instruct the child agent to return a short plain-text response without Markdown tables, citations, or multiple sections. This will help rule out response payload size or channel-rendering issues.
    • Test the same questions directly against Agent B in Microsoft 365 Copilot, if possible. This can isolate whether the failure is specific to the parent-to-child handoff.
    • Republish both agents, then remove and reconnect Agent B to Agent A to refresh the connected-agent reference.
    • Confirm that both agents use compatible authentication settings and that no additional connected-agent chaining is occurring.
    • Enable Application Insights logging for incoming and outgoing activities and capture the conversation ID, timestamp, activity ID, and any error code generated during the failed turn.

    Copilot Studio documents errors such as ConversationStateTooLarge, AsyncResponsePayloadTooLarge, connected-agent authentication errors, execution timeouts, and Dataverse search throttling. However, based on the described behavior, conversation-state growth or a Microsoft 365 channel handoff issue seems more likely than SharePoint retrieval failure.

    Reference Document:
    1.https://learn.microsoft.com/en-us/troubleshoot/power-platform/copilot-studio/authoring/error-codes?tabs=webApp
    2.https://learn.microsoft.com/en-us/microsoft-copilot-studio/advanced-bot-framework-composer-capture-telemetry
    3.https://learn.microsoft.com/en-us/microsoft-copilot-studio/telemetry-overview
    4.https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-transcript-controls
    5.https://learn.microsoft.com/en-us/microsoft-copilot-studio/analytics-transcripts-studio

    Was this answer helpful?

    0 comments

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.