web
You’re offline. This is a read only version of the page.
close
Skip to main content

Announcements

Welcome to the Power Platform Communities
The Next Big Bite | Aussie Classics & Starting Small with Clay Wesener | EP04
News and Announcements icon
Community site session details

Community site session details

Session Id :
Power Platform Community / Forums / Copilot Studio / Several Issues-Buildin...
Copilot Studio
Suggested Answer

Several Issues-Building a Copilot Studio Agent

(0) ShareShare
ReportReport
Posted on by HP-16060428-0 4

Hello,

I am currently building a Main Agent with a Connected Child Agent in Microsoft Copilot Studio and have encountered several issues. I would appreciate your guidance on whether these are expected behaviors or if there are recommended configurations to resolve them.

Environment

  • Main Agent with uploaded Knowledge
  • Connected Child Agent (GS Corporate Document Agent)
  • Child Agent is referenced using the /Agent syntax in the Main Agent instructions
  • Both Main Agent and Child Agent are published
  • The issue occurs in both Copilot Studio Test Chat and Microsoft Teams

Issue 1: Duplicate responses after Connected Agent routing

When a request is routed to the Child Agent, the Child Agent responds correctly.

However, after the Child Agent finishes, the Main Agent generates an additional response instead of stopping.

For example:

User: 공장등록증 줘

Expected behavior:

  • Child Agent returns the predefined Topic response once.

Actual behavior:

  • Child Agent returns the Topic response.
  • Child Agent returns the document folder link.
  • Then the Main Agent generates another response (or fallback), resulting in duplicate answers.

Is there a supported way to ensure that the Main Agent completely stops processing after the Connected Agent returns its response?

Issue 2: Incorrect response language

The Main Agent is configured with the following instruction:

  • Respond in the same language as the user's question.
  • Korean questions → Korean responses only.
  • English questions → English responses only.

However, after the Child Agent responds, the Main Agent sometimes generates an additional response in English or mixes Korean and English, even though the original question was asked in Korean.

How can I ensure that the response language always matches the user's language, including after Connected Agent routing?

Issue 3: Topic priority over Knowledge

My intended behavior is:

  1. Check Topic first.
  2. If no Topic matches, search Knowledge.
  3. If no Knowledge is found, return the fallback response.

However, in my testing, the Main Agent often does not follow this order consistently.

I originally tried implementing company document requests (such as Business Registration Certificate and Factory Registration Certificate) using Knowledge. However, the agent frequently failed to find the relevant Knowledge and returned "No information found," even though the content existed.

Because of this, I moved those document requests into a Connected Child Agent with predefined Topics.

Is there a recommended way to enforce the following execution order?

  1. Always prioritize Topic.
  2. Only search Knowledge if no Topic matches.
  3. Execute the fallback response only if neither Topic nor Knowledge provides an answer.

Could you please advise whether these behaviors are expected with Connected Agents and Generative Orchestration, or if there are recommended settings or best practices to achieve the expected behavior?

Thank you in advance for your support.Hello,

I am currently building a Main Agent with a Connected Child Agent in Microsoft Copilot Studio and have encountered several issues. I would appreciate your guidance on whether these are expected behaviors or if there are recommended configurations to resolve them.

Environment

  • Main Agent with uploaded Knowledge
  • Connected Child Agent (GS Corporate Document Agent)
  • Child Agent is referenced using the /Agent syntax in the Main Agent instructions
  • Both Main Agent and Child Agent are published
  • The issue occurs in both Copilot Studio Test Chat and Microsoft Teams

Issue 1: Duplicate responses after Connected Agent routing

When a request is routed to the Child Agent, the Child Agent responds correctly.

However, after the Child Agent finishes, the Main Agent generates an additional response instead of stopping.

For example:

User: 공장등록증 줘

Expected behavior:

  • Child Agent returns the predefined Topic response once.

Actual behavior:

  • Child Agent returns the Topic response.
  • Child Agent returns the document folder link.
  • Then the Main Agent generates another response (or fallback), resulting in duplicate answers.

Is there a supported way to ensure that the Main Agent completely stops processing after the Connected Agent returns its response?

Issue 2: Incorrect response language

The Main Agent is configured with the following instruction:

  • Respond in the same language as the user's question.
  • Korean questions → Korean responses only.
  • English questions → English responses only.

However, after the Child Agent responds, the Main Agent sometimes generates an additional response in English or mixes Korean and English, even though the original question was asked in Korean.

How can I ensure that the response language always matches the user's language, including after Connected Agent routing?

Issue 3: Topic priority over Knowledge

My intended behavior is:

  1. Check Topic first.
  2. If no Topic matches, search Knowledge.
  3. If no Knowledge is found, return the fallback response.

However, in my testing, the Main Agent often does not follow this order consistently.

I originally tried implementing company document requests (such as Business Registration Certificate and Factory Registration Certificate) using Knowledge. However, the agent frequently failed to find the relevant Knowledge and returned "No information found," even though the content existed.

Because of this, I moved those document requests into a Connected Child Agent with predefined Topics.

Is there a recommended way to enforce the following execution order?

  1. Always prioritize Topic.
  2. Only search Knowledge if no Topic matches.
  3. Execute the fallback response only if neither Topic nor Knowledge provides an answer.

Could you please advise whether these behaviors are expected with Connected Agents and Generative Orchestration, or if there are recommended settings or best practices to achieve the expected behavior?

Thank you in advance for your support.

Categories:
Autonomous agents
I have the same question (0)
  • Suggested answer
    Sam_Fawzi Profile Picture
    Sam_Fawzi 957 Super User 2026 Season 1 on at
    Hey @ HP-16060428-0  
     
    The three symptoms mostly come from one root cause: in generative orchestration the parent isn't a router that forwards the child's answer, it's an LLM planner that gathers from the child and then generates its own final response on top. So you get the child speaking plus the parent re-generating.
     
    1. Duplicate responses. The child is replying to the user directly (Topic message + link), then the parent adds another pass. Per Microsoft's multi-agent guidance, subagents don't know they're subagents and will reply to the user by default, causing duplicate messages. Fixes: tell the child in its instructions that it's a subagent returning content for the parent to relay; return the child's answer as a topic output variable instead of a Message node; and tell the parent to relay verbatim without adding a second answer. For canned document links, consider an inline topic instead of a connected agent, connected agents each run their own orchestration, which is what adds the extra pass.
    2. Language drift. This is a side effect of #1 — the parent's extra generation is a separate call that drops the language rule. Put the "match user language" instruction in the child too, and fix #1 so there's only one generation point to control.
    3. Topic | Knowledge | Fallback order. Honestly, that strict order doesn't exist in generative orchestration. The planner picks based on names/descriptions and you can't force it. If you need deterministic ordering, use classic orchestration,  topics fire on trigger phrases first, knowledge sits in Conversational Boosting as fallback. If you stay generative, you can only bias it with very specific, non-overlapping topic descriptions. Moving exact document lookups out of knowledge into topics is the right call; keep knowledge for open-ended Q&A.
    References:
  • HP-16060428-0 Profile Picture
    HP-16060428-0 4 on at

    I understand why duplicate responses occur with Connected Agents. However, my reason for creating a separate Child Agent was that when Topics and Knowledge coexist in the same agent, the Topic always takes priority and the Knowledge is never used.

     

    If I move the document links back into Topics as suggested, I'll run into the same issue again.

     

    In Generative Orchestration, what architecture do you recommend if I need both:

     


    1. Knowledge-based Q&A (RAG) for manuals, and


    2. Deterministic document links for specific keywords,


      3.  
     

    without duplicate responses and without losing the ability to search Knowledge?

     

    Is there a supported architecture for this scenario?

  • Sam_Fawzi Profile Picture
    Sam_Fawzi 957 Super User 2026 Season 1 on at
    Great follow-up, this gets to the heart of it. The short answer is: yes, there's a supported architecture for this, and the good news is you can do it in a single agent without the duplicate-response problem.
    The key thing to understand is that generative orchestration has no built-in "topics beat knowledge" rule. Routing is entirely description-driven,  the planner reads the name and description of every topic, tool, and knowledge source to decide what to call. So the behavior you saw ("topic always wins, knowledge never used") wasn't a priority setting you were fighting. It was a broad topic description quietly capturing queries that should have gone to knowledge. That's fixable, and fixing it lets both paths coexist.
     
    Recommended architecture (one agent, generative orchestration):
     
    1. Manuals + knowledge sources at the agent level. This stays your default RAG path and answers anything not explicitly claimed by a tool.
     
    2. Deterministic document links + a Tool/Action rather than a Topic. Give it a parameterized input (e.g. document name) and a description that names exactly which documents/keywords it serves. An action with a clear input contract routes more reliably than a topic, because the planner treats it like a callable function instead of something competing with knowledge for open-ended questions.
     
    3. Add negative scoping to your descriptions (this is the lever that solves the starvation). Microsoft's guidance is explicit that you should use the description to say when an item should not be used. For example, on the document tool: "Use only when the user requests a specific named certificate or document. Do NOT use for how-to, procedure, or manual questions, those are answered from knowledge." Mutually exclusive descriptions are what keep the two paths from cannibalizing each other.
     
    4. No child agent. With everything in one agent there's no second orchestration pass, so the duplicate responses go away as a side effect.
    If you ever need to pin a specific knowledge source to a specific intent, you can embed that source in a generative answers node inside a topic, otherwise the agent searches all agent-level knowledge.
     
    One honest caveat: routing stays probabilistic,  you can't force the agent to use a specific knowledge article on demand. Use the activity map and Get rationale in the test panel to see which resource the planner chose per query, and tighten descriptions until it's consistent. If you truly need guaranteed deterministic ordering with zero variance, that only exists in classic orchestration, but you'd give up multi-intent Q&A ,  so for your dual requirement, the single-agent + action + scoped-descriptions pattern is the right fit.
     
    References:
    Configure high-quality instructions (when-not-to-use guidance): https://learn.microsoft.com/en-us/microsoft-copilot-studio/guidance/generative-mode-guidance
    Apply generative orchestration (description-driven routing, can't force knowledge): https://learn.microsoft.com/en-us/microsoft-copilot-studio/guidance/generative-orchestration
    Knowledge sources (scoping into a generative answers node): https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio

Under review

Thank you for your reply! To ensure a great experience for everyone, your content is awaiting approval by our Community Managers. Please check back later.

Helpful resources

News and Announcements

Welcome to the Power Platform Communities
The Next Big Bite | Aussie Classics & Starting Small with Clay Wesener | EP04

Quick Links

Season of Sharing Community Challenge Winners!

Congratulations to our community stars!

Kudos to our 2025 Community Spotlight Honorees

Expanding mentorship, skilling, and AI innovation

Congratulations to the June Top 10 Community Leaders!

These are the community rock stars!

Leaderboard > Copilot Studio

#1
sannavajjala87 Profile Picture

sannavajjala87 160 Super User 2026 Season 1

#2
11manish Profile Picture

11manish 145

#3
Haque Profile Picture

Haque 121

Last 30 days Overall leaderboard

Featured topics

Announcing the "Microsoft Copilot Studio ❤️ MCP" lab
Block PII/PCI in Copilot Studio agent user prompt

Product updates

Microsoft Power Platform Community release plans