Thank you for the clarification — this makes things much clearer. The issue is not related to the firewall or payload size, but rather a rendering bug specific to the Copilot Studio ↔ Foundry integration.
The symptom you describe is precise: the response arrives, is visible in the evaluations, but the chat stays stuck on the progress indicator. This is a known behavior when the connection between Copilot Studio and the Foundry agent enters an inconsistent state.
Checklist to follow in order
1. Verify that the Foundry agent was created in the new Microsoft Foundry portal
You can only connect to agents created in the new Microsoft Foundry portal. Connecting to an agent created in the previous portal generates a 404 - Version not found error. Even if you don't see an explicit error, a connection to an agent from the old portal can cause silent rendering failures. Microsoft Learn
2. Make sure the connected agent has Description and Instructions filled in
The connected agent must have a clear Description and Instructions configured on its Overview page. After making changes, you need to publish the connected agent and then return to the orchestrator agent to test again. Microsoft Learn
If the Description field of the Foundry agent in Copilot Studio is empty or too generic, the orchestrator may select it correctly (as you see in evaluations) but then fail to render the response due to missing GPT metadata.
3. Disconnect and reconnect the Foundry connection — but with a variation
You mentioned it worked temporarily after a disconnect/reconnect. This is the classic signal of a stale token connection. The stable workaround is:
- Go to Settings → Connections in your Copilot Studio agent
- Completely remove the connection to the Foundry agent
- Publish the Foundry agent directly from the Foundry portal (make sure it's on an active stable version)
- Recreate the connection in Copilot Studio
- Publish the orchestrator before testing — after creating the connection, you must provide an appropriate Name and Description for the SDK agent, so the orchestrator knows when to use it. Microsoft Learn
4. Verify the active published version in the Foundry portal
What gets published is the agent's stable endpoint — end users always interact with a consistent agent entity. To update the active version, you need to select the correct version in the Foundry portal using the version selector. If the Foundry agent doesn't have a correctly published active version, Copilot Studio receives the response but fails to resolve it for rendering. Microsoft Learn
If none of the above resolves the issue
Since the problem reproduces consistently even with the simple chit-chat agent (no tools or knowledge), this is most likely a platform-level bug in the Copilot Studio ↔ Foundry integration. In that case:
- Open a Microsoft Support ticket bringing the Session IDs visible in your evaluations — these are critical for the engineering team to trace the issue
- Include the screenshots you already have (evaluation showing response received + chat not displaying anything)
- Specify that the issue reproduces even with the chit-chat agent without tools or knowledge, ruling out payload or permission-related causes
This bug profile (response received but not rendered, temporarily fixed by reconnecting) is exactly the type of regression Microsoft resolves via hotfix once they have the Session IDs to trace it.
Hope this helps!
Paolo
✅ Did this solve your issue? → Accept as Solution
👍 Partially helpful? → Click "Yes" on "Was this reply helpful?" or drop a Like!
Want more tips on Power Platform & AI? Follow me here:
🔗 LinkedIn: https://www.linkedin.com/in/paoloasnaghi/
▶️ YouTube: https://www.youtube.com/@BeyondThePlatforms
📸 Instagram: https://www.instagram.com/beyond_the_platforms/
🌐 Website: https://www.beyondtheplatforms.com/
