Hey everyone,
I'm running into a frustrating set of issues with a multi-agent setup in Microsoft Copilot Studio and wanted to see if anyone else has hit these or found workarounds.
My Setup
- Orchestrator / Main Agent
- Connected Agents (fully published, connected to the main agent)
- Channels: MS Teams/ Copilot studio test
Issues I'm Seeing
1. Conversation Boosting Topic firing during multi-turn conversations
Even when the conversation history clearly contains context for a follow-up question (i.e., the user is mid-conversation and asking a clarifying question), the orchestrator keeps falling back to the Conversation Boosting system topic instead of continuing the active topic or delegating to the appropriate connected agent. It looks like the LLM-based orchestration isn't properly using conversation history to maintain context.
2. Connected Agent not getting triggered
Some of the time, even when the user's intent clearly maps to what one of my connected agents is designed to handle, the orchestrator doesn't invoke that connected agent at all – it either tries to answer itself or falls back to Conversation Boosting again. My connected agents are published and the agent instructions in the orchestrator do describe when to hand off to each connected agent.
3. Escalation topic triggering unexpectedly
The Escalation system topic is triggering far too often – in situations where the query should have been handled either by the orchestrator or a connected agent. This is happening even on straightforward questions.
Key Observation – Model-Specific Behaviour
This is the most important clue I've found so far: all three issues above only occur when the orchestrator is using GPT-5-chat. When I switch the orchestrator model to GPT-4.1, the behaviour is stable – conversation boosting doesn't fire inappropriately, connected agents get triggered correctly, and escalations are rare.
This strongly suggests the issue is related to how the newer model interprets orchestrator instructions and conversation history in a multi-agent context, rather than a configuration problem on my end. Has anyone else observed this model-specific behaviour?
What I've Already Tried
- Verified connected agents are published before being connected to the main agent
- Reviewed agent instructions on the orchestrator to ensure clear handoff conditions are described
- Checked that conversation history passing is enabled between orchestrator and connected agents
- Tested connected agents in isolation – they work fine independently
- Switched orchestrator model from GPT-5-chat → GPT-4.1 → issues disappear
My Questions
- Is there a known compatibility issue between the newer GPT model GPT-5-chat and multi-agent orchestration in Copilot Studio?
- Does the newer model require different instruction patterns or prompt formatting to reliably delegate to connected agents across multi-turn conversations?
- Is there a way to suppress or deprioritize system topics (Conversation Boosting, Escalation) so the LLM orchestration layer takes precedence?
- Is this a platform-level bug that Microsoft is tracking? Worth raising a support ticket?
Environment Info
- Copilot Studio version: [Latest as of your date]
- Agent type: Connected Agents (not child agents)
- Orchestration mode: Generative (AI-based orchestration)
- Model where issue occurs: GPT-5-chat
- Model where it works fine: GPT-4.1
Any help, documentation pointers, or workarounds would be hugely appreciated. For now I've rolled back to GPT-4.1 on the orchestrator, but I'd love to use the newer model without these instabilities.
Thanks!
