Copilot agent is giving some garbage value

(0) Share

Report

Posted on by VR-18030804-0 Microsoft Employee

We have a Copilot Studio parent agent that is connected to multiple Microsoft Fabric Data Agents and is published to Microsoft Teams. When queries are executed directly against the individual Fabric Data Agents (as shown in the attached screenshots), the agents return correct and validated results, confirming that the data sources, KQL generation, and Fabric agent logic are working as expected. However, when the same queries are asked through the Copilot agent in Microsoft Teams, the responses are inconsistent and often incorrect, sometimes returning partial, irrelevant, or “garbage” values. The issue is reproducible with simple queries such as row count checks, where the Fabric agent correctly returns the expected row count, but the Teams Copilot response does not match that output. The Copilot Studio configuration shows all child agents are properly connected, enabled, and published, with no visible configuration errors. This indicates that the problem is not with the Fabric Data Agents themselves, but likely with the Copilot Studio orchestration or Teams runtime handling of agent responses, such as context passing, state management, or response grounding. Assistance is needed to debug why the Teams channel is not reliably reflecting the Fabric agent outputs and to ensure consistent behavior across Fabric, Copilot Studio testing, and Microsoft Teams.

Could someone please assist us in debugging this issue with the Copilot Studio agent behavior in Microsoft Teams?

Categories:

Copilot Studio skills development

Calling actions from Copilot Studio

I have the same question (0)

All responses (4)

Answers (0)

Sort by

Suggested answer

Nivedipa-MSFT Microsoft Employee on at

Like
a
(0)

Report
Copy link

Link copied!
Hello VR-18030804-0
The Fabric agents provide accurate data directly, but the parent Copilot agent in Teams is returning incorrect values. This indicates the issue lies with parent agent orchestration rather than the data agents themselves.

Possible causes include:

The parent agent is routing to the wrong child agent, especially if multiple agents have overlapping data scopes.

The parent’s generative AI may be modifying or regenerating numeric values instead of passing them through exactly as received.

Recommended solutions:

Trace routing using the Copilot Studio test pane and the "Track between topics" feature to verify which child agent is invoked.

Refine child agent descriptions to ensure they are distinct and avoid overlap, helping the orchestrator select the correct agent.

Implement explicit routing topics by creating topics with trigger phrases that direct queries to specific child agents, rather than relying solely on generative orchestration.

Add clear instructions to the parent agent, such as: "For data queries, relay the child agent’s exact response without modification. Never approximate or regenerate numeric values."

Was this reply helpful? Yes No
LY-18031556-0 Microsoft Employee on at

Like
a
(0)

Report
Copy link

Link copied!

Hi Nivedipa,

We have already followed the steps you have given. Our child agents work fine and never give garbage data. Our orchestrator always correctly routes to the child agent. We have given the orchestrator agent explicit instructions to always delegate and return the answer from the child agent verbatim. Nevertheless, the orchestrator agent sometimes produces garbage data.

Please advise.

Was this reply helpful? Yes No
LY-18031556-0 Microsoft Employee on at

Like
a
(0)

Report
Copy link

Link copied!

Hi Nivedipa,

Following up on this thread. We are still experiencing issues, and we have already implemented all of the practices mentioned in the suggested answer.

Was this reply helpful? Yes No
Nivedipa-MSFT Microsoft Employee on at

Like
a
(0)

Report
Copy link

Link copied!
Hello LY-18031556-0,
This confirms the issue is within the Copilot Studio orchestration runtime, not the configuration. Here are some recommended next steps:

1. Disable generative orchestration for data topics

The generative AI layer is overriding your instructions and reinterpreting child agent responses. To avoid this:

Create explicit topics for each child agent within the parent agent, rather than relying on generative orchestration.

Within each topic, use the "Transfer to agent" action, followed by a "Send a message" node that uses the Activity.Text variable from the child agent.

This will bypass the generative synthesis layer entirely.

2. Use plugins or actions instead of agent-to-agent connections

Instead of connecting Fabric agents as child agents:

Expose the Fabric Data Agent as a Copilot Studio plugin action.

Invoke it through a Power Automate flow from the parent agent.

Return the flow output directly to have full control over the response.

3. Investigate possible Teams-specific rendering issues

Check if the incorrect data also appears in the Copilot Studio test pane.

If the issue is only in Teams, the channel adapter could be affecting the response. Try re-publishing to Teams and testing in a new chat.

If it appears in the test pane as well, the orchestration runtime is likely the cause.

4. Enable Application Insights tracing

If not already done:

Navigate to Copilot Studio > Settings > Advanced > Application Insights.

Connect to an Application Insights resource.

Run this query:

customEvents
| where name contains "GenerativeOrchestration" or name contains "ChildAgent"
| project timestamp, name, customDimensions
| order by timestamp desc

Compare the child agent response payload with the final orchestrator output to see where changes occur.

5. Consider a potential platform bug

If the Activity Trace shows the child agent provides correct data but the orchestrator still produces incorrect results, even with explicit instructions, this likely points to a Copilot Studio orchestration bug. In this case, you can share the below data to raise a bug:

Screenshots of Activity Trace showing the correct child agent output and the incorrect final response

Your parent agent instructions, including the pass-through directive

Specific queries with both expected and actual outputs

Was this reply helpful? Yes No