Hi, I’ve created a basic agent for enquiries using copilot studio to to connect to BC MCP server and then deploying via teams.
While the system works, it’s quite fragile. It repeatedly gives authentication errors with no way to resolve without either just waiting or republishing. Also any “larger” or less defined query “find highest margin item sold last week” seems to hang and crash the system.
Is this a Copilot studio issue or some kind of configuration? I find the interface useful, but can’t deploy with this type of stability.
I’ve done a similar test with Claude and registering in Azure, and better - but didn’t get around to deploying in Teams yet. I’d like to build a suite of agent and find teams is a good place for them to “live”.
I realize this topic is changing rapidly but I don’t want to change platform\structure repeatedly. Maybe Copilot Studio will improve in next couple of months? Any thoughts or suggestions welcome.
Hi @Kamal_Shah
Sounds like you’re hitting real stability issues. Authentication problems and hanging on complex queries is common with this kind of setup right now.
Check your connection settings and timeouts, but honestly, these are known issues.
Complex questions don’t work as well. Keep it simple and specific. Your Claude test did better because it handles complicated stuff more gracefully.
Real talk: If you need it working now, the platform isn’t quite there yet. If you can wait a few months, it might improve. But if stability matters, use what works better right now.
Maybe start simple first, just basic questions and lookups. See how it goes. Then decide if it’s good enough for what you need.
We’ve seen similar behavior in early MCP and Copilot Studio implementations. Authentication issues are often related to token expiration, session handling, or Teams/Copilot connection refresh rather than the BC MCP server itself.
For larger, less-defined queries such as “highest margin item sold last week,” the issue may be caused by the agent generating inefficient tool calls, returning large datasets, or hitting timeout and context limits. Try narrowing the query scope and monitoring the MCP server logs to identify where the request is failing.
If stability is a priority, I would be cautious about deploying business-critical solutions on the current MCP/Copilot stack without extensive testing. The platform is evolving rapidly, and reliability is improving, but there are still edge cases around authentication, tool orchestration, and long-running queries.
Have you checked the Copilot Studio analytics, Application Insights logs, or BC telemetry to see whether the failures occur during authentication, tool execution, or response generation? That may help isolate whether this is a configuration issue or a platform limitation.