Child agent works standalone but fails when invoked via orchestrator

(1) Share

Report

Posted on by SR-24020711-0

Hi Community,

I’m working in Microsoft Copilot Studio with a Parent/Orchestrator agent that routes queries to 4 Child agents.

For 3 out of 4 child agents, everything works as expected:

When I ask a child-agent-related question to the Orchestrator, it routes to the correct child agent and returns the same (correct) response as when testing the Child agent directly.

However, for one specific child agent, I’m seeing inconsistent behavior:

What I’m observing

If I ask the question directly to the Child agent, the Child agent:
- triggers the correct topic
- executes the expected flows/tools
- returns the correct answer
If I ask the same question to the Parent/Orchestrator, the Orchestrator:
- does route/redirect to that same Child agent
- but the Child agent does not follow its expected topics and flows
- and the response becomes incorrect / falls back / fails to fetch the intended output

So routing seems correct, but the Child agent behavior changes when invoked via the Orchestrator.

What I’ve already validated:

The child agent works correctly in isolation.
The same prompt/question is used in both tests.
Orchestrator routes to the correct child agent (routing is not the issue).

Any guidance, best practices, or troubleshooting steps would be greatly appreciated. Thanks in advance!

Categories:

Copilot Studio pre-built agents/templates

I have the same question (0)

All responses (3)

Answers (1)

Sort by

André Arnaud de Cal... 545 on at

Like (2)

Report
Copy link

Link copied!

Hi,

Can you share more details about the description and instructions used in your parent agent? In case you are using AI based orchestration, the prompt used by the agent to call the child might be different, resulting in a different outcome.

Was this reply helpful? Yes No
Suggested answer

Valantis 3,470 on at

Like (2)

Report
Copy link

Link copied!

Hi @SR-24020711-0,

This happens because when the orchestrator calls the child agent it passes the full conversation history, which changes how the child interprets the query compared to a clean direct test. The child also does not trigger its topics the same way when invoked via a parent.

Before changing anything, open the activity map in the orchestrator test pane and reproduce the issue. It shows exactly what the child receives and what it picks internally. That will tell you where it breaks.

Then check these by order:

1. Tighten the child agent description in the orchestrator. If the description is vague the child picks the wrong path.
2. Check if you are passing too much conversation history. Explicitly pass only what the child needs.
3. In the child agent settings configure the Outputs, and in the parent set the After running behavior. Without this the orchestrator can ignore the child response and generate its own fallback.
4. If still wrong, temporarily switch that child to classic orchestration and test. If it works there the issue is the generative orchestration picking the wrong topic.

Ref: https://learn.microsoft.com/en-us/microsoft-copilot-studio/guidance/multi-agent-patterns

Best regards,

Valantis

✅ If this helped solve your issue, please Accept as Solution so others can find it quickly.

❤️ If it didn’t fully solve it but was still useful, please click “Yes” on “Was this reply helpful?” or leave a Like :).

🏷️ For follow-ups @Valantis.

📝 https://valantisond365.com/

Was this reply helpful? Yes No
Verified answer

SR-24020711-0 26 on at

Like (2)

Report
Copy link

Link copied!
Issue resolution steps:

Child Agent - Removed the connection reference in settings

Published the child Agent (Important Step)

Rerun the query on Orchestrator.

Was this reply helpful? Yes No