Issues with Multi-Agent Orchestration in Microsoft Copilot Studio (Teams vs. Test Behavior)

(2) Share

Report

Posted on by 菜日-20110828-0

I’m building a multi-agent solution in Microsoft Copilot Studio with a parent agent acting as a “Reception” bot and three child agents (System QA Bot, HR Agent, and General Affairs Agent). The parent uses generative orchestration to route queries to the appropriate child agent based on intent.

Problem:

In the Copilot Studio test environment, everything works as expected. For example, when I ask “I want to create an email signature,” the parent correctly calls the System QA Bot, and the child agent responds using its knowledge base.

However, when I publish the parent agent and run it in Microsoft Teams, the behavior changes:

The parent correctly calls the child agent.

The child agent immediately falls back (even for simple queries that work fine in test).

After the fallback, the parent agent generates its own answer using orchestration, which is often irrelevant.

What I’ve confirmed:

Generative orchestration is enabled only on the parent agent (as required for multi-agent).

Child agents are published and work perfectly when used standalone in Teams.

“Pass conversation history” is enabled.

Parent agent has zero knowledge sources (it’s meant to be a pure router).

The fallback message from the child agent is customized, and I see that message in Teams, so the child is definitely being invoked.

Observation:

When checking the child agent’s activity logs, the same query produces very different results:

In the test environment, the child retrieves relevant knowledge and generates a detailed answer.

In Teams, the same query shows a 0% confidence score and no knowledge retrieved.

Question:

Has anyone else experienced this discrepancy between test and Teams when using multi-agent orchestration?

Why would the child agent fail only when invoked via the parent agent in Teams?

Is there any way to adjust fallback timing or confidence thresholds?

Are there best practices to prevent the parent from generating answers after a child fallback?

Any insights or workarounds would be greatly appreciated!

Categories:

Building Copilot Studio chatbots in Microsoft Teams

I have the same question (3)

All responses (3)

Answers (0)

Sort by

Romain The Low-Code... 2,876 Super User 2026 Season 1 on at

Like
a
(0)

Report
Copy link

Link copied!

hello, thanks it's interesting,

i have nothing to help , just to say, it's nearly the third or fourth time i see this kind of post here :) i hope Microsoft will make something for this matter

Was this reply helpful? Yes No
NM-21100743-0 2 on at

Like
a
(0)

Report
Copy link

Link copied!

Hi

Have you tried adding an output in the Completion section of your child agent. I had something similar where everything looked great running the tests in the Studio however as soon as I launched in Teams or 365 "Sorry try rephrasing......". I found once I added an output my agents behaved as I expected when querying the sub agents knowledge.

Hope this helps.

Was this reply helpful? Yes No
Suggested answer

Sayali-MSFT Microsoft Employee on at

Like
a
(0)

Report
Copy link

Link copied!
Hello,
Child agents work in test mode because full context and orchestration metadata are shared. In Teams, handoff often passes incomplete context or stripped metadata, causing the child agent’s confidence to drop to 0% and trigger fallback—often due to channel connector limits or production permission issues accessing knowledge sources.
When this happens, the parent agent detects the fallback as “no answer” and generates its own response.

Key Fixes

Explicitly pass full context and handoff metadata

Verify child agent knowledge source access in Teams

Lower the child agent’s confidence threshold

Prevent the parent from generating fallback responses

Compare test vs Teams logs for missing or altered context

Workarounds

Add a clarifying parent message before handoff

Invoke child agents directly via Teams deep links

Was this reply helpful? Yes No