Generative AI Orchestration degraded/ working inconsistently

(1) Share

Report

Posted on by CU17030703-1

Hello,

I am using gpt-5 in Copilot Studio. My setup has a main agent that understands and clarifies user requirements and uses a child agent to search Sharepoint documents to answer user query. The agent is published and accessed as a Teams app. This setup was working fine till last week. Yesterday, we noticed that Teams app is frequently using the Fallback 'I'm sorry, I'm not sure how to help with that. Can you try rephrasing?'. It does reply once after a startover or reset, but then falls back to rephrase message for subsequent queries

When debugging in Copilot Studio, it is returning rephrase for all scenarios. Looking at the behaviour across many scenarios, it appears that the behaviour of Generative AI Orchestration has degraded. It is not able to ask clarifying questions and not able to invoke child agent for searching Sharepoint docs. When Sharepoint docs are connected directly to main agent, it starts searching them, but still does not ask clarifying questions when context is not clear.

Any pointers on why the behaviour might have changed? Any known open issues in Microsoft Studio that could have led to this issue or any latest releases that were pushed recently that could have caused this?

Any pointers on debugging the reasoning of Generative AI Orchestrator would be useful as well. In my case, I see only a start node when it gets into rephrase message mode. So not able to understand why it doesn't ask clarifying questions or doesn't route to child agent.

Thanks

Categories:

General topics

I have the same question (0)

All responses (6)

Answers (0)

Nivedipa-MSFT Microsoft Employee on at

Like (0)

Report
Copy link

Link copied!

Thanks for bringing this issue to our attention. We will check and update you soon.

Was this reply helpful? Yes No
CU17030703-1 4 on at

Like (0)

Report
Copy link

Link copied!

Thanks for looking into this. For additional context, in my setup, Use general knowledge under Settings > Generative AI is disabled. Enabling this setting lead to agent asking clarification questions and routing to the child agent. In some cases, it leads to agent answering questions from its own knowledge, and hence we would like it keep it disabled, once our child agent routing starts working again.

Was this reply helpful? Yes No
Suggested answer

Nivedipa-MSFT Microsoft Employee on at

Like (0)

Report
Copy link

Link copied!
Hello CU17030703-1,
Turning off "Use general knowledge" will now also prevent the orchestrator from reasoning, classifying intent, asking clarifying questions, and routing to child agents. This change is due to a recent platform update that more closely links these features, whereas they previously operated separately.

Workaround (Keep General Knowledge Off)

Override the Fallback topic — direct all unmatched queries to your child agent instead of displaying "rephrase"

Include explicit trigger phrases for each topic (5-10 variations) — this removes the need for generative intent classification

Manually add clarifying questions using Question nodes in your routing topics

Republish after making these updates

Was this reply helpful? Yes No
CU18031017-0 4 Microsoft Employee on at

Like (0)

Report
Copy link

Link copied!

Thanks for your response. This is really helpful. However, this is a major behaviour change. It would be a common pattern to have the main agent route to child agents for domain specific knowledge sources. For instance, child agents could answer queries specific to HR, Finance, and Sales departments from their respective knowledge bases. As you can imagine, most of these queries can also be answered from LLM's general knowledge. Hence, it is critical to have the responses grounded in company specific resources.

Are there any release notes that I can refer that mention the specific change and the rationale?

Considering, many users may be impacted by this change, can we please raise an internal ticket to provide a clean workaround for enabling this parent-child pattern?

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

Like (0)

Report
Copy link

Link copied!

Hello CU18031017-0 - We have highlighted this issue to the engineering team and are raising a bug to investigate the recent degradation in Generative AI Orchestration behavior, especially in multi‑agent setups leveraging child agents and SharePoint knowledge sources.

Was this reply helpful? Yes No
CU18031017-0 4 Microsoft Employee on at

Like (0)

Report
Copy link

Link copied!

Hello @Nivedipa-MSFT, are there any updates on this issue from the engineering team?

Was this reply helpful? Yes No