Foundry Agent responds in Playground but silent in Web Chat and Teams — Activity Protocol response not reaching channel

Question

I have an agent built in the new Microsoft Foundry (non-hub-based) using gpt-4o with several custom tools (ManageMicrosoft365Account, custom utilities, File Search, and Grounding with Bing Search). The agent works perfectly in the Foundry Playground — responds to prompts, calls tools, returns results.

When I publish to Teams and Web Chat using the built-in "Publish to Teams and Microsoft 365 Copilot" flow, the agent does not respond. The chat window stays completely silent with no error message.

Here's what I've confirmed through troubleshooting:

• The initial Bot Service showed HTTP 401 Unauthorized errors in the Web Chat channel logs. I deleted it and republished from Foundry, which created a new Bot Service with auto-configured Activity Protocol endpoint.
• All channels (Direct Line, Microsoft Teams, Web Chat) now show as Healthy — no more auth errors.
• Azure AI User role has been assigned to the project managed identity, Foundry resource identity, hub identity, and Bot Service principal on the Foundry resource.
• The Foundry Monitor tab confirms agent runs are being triggered (41 started, 28 completed, 10 failed).
• Traces confirm the agent receives input and successfully generates output. For example, input "Hi" produced output "Hi, Matt! What's on your mind?" — visible in the Traces tab.
• However, that response never appears in Web Chat or Teams.
• A subsequent run showed the agent received input but generated no output at all.
• Tested with and without Enable Streaming Endpoint — no change.
• Even a simple "Hi" prompt with no tool invocation fails to return a response.

Conclusion: The agent executes successfully in Foundry, but the Activity Protocol response path from the Agent Application back through the Bot Service to the channel is silently dropping the response.

Has anyone successfully deployed a Foundry agent with tools to Teams using the one-click publish flow? Is there a known workaround, or is this a current platform limitation with the Activity Protocol relay?

Environment: New Foundry (non-hub), Single Tenant Bot, Activity Protocol endpoint auto-configured by publish flow.

Answer 1

Hi Matt Krudop (SFNexa),

Based on similar issues we’ve seen, when an agent responds correctly in the Foundry Playground but stays silent in Teams or Web Chat, the problem is usually not the agent logic, but the delivery path back to the channel.

Here are some concrete things to check and try:

1. Confirm the agent is actually producing output

You’ve already done this, but it’s important:

• Go to Foundry → Monitor / Traces
• Verify that each Teams/Web Chat message:
  • Triggers an agent run
    • Shows a generated response text in the trace

If output exists in Traces but never reaches the channel, this confirms the issue is in the Activity Protocol relay, not the model or tools.

2. Test with the simplest possible agent response

To rule out tool‑related or payload issues:

• Temporarily disable all tools
• Test with a trivial response like:

User: Hi
Agent: Hi there

If even this doesn’t show up in Teams/Web Chat, the issue is not caused by tools, grounding, or MCP, and is almost certainly platform‑side.

3. Toggle Streaming (but don’t expect it to fix it)

Some users tested both:

• Enable Streaming Endpoint = ON
• Enable Streaming Endpoint = OFF

In reported cases, this did not change the behavior, but it’s still worth validating so you can tell support you tested both paths.

4. Re‑publish to create a fresh Bot Service

If the bot was previously showing auth errors:

• Delete the existing Azure Bot Service
• Re‑publish the agent from Foundry using the one‑click publish flow
• Confirm all channels (Teams, Web Chat, Direct Line) show Healthy

This rules out stale or misconfigured Activity Protocol endpoints.

5. Double‑check Managed Identity assignments (again)

Even if this was already done, re‑verify that nothing regressed:

• Azure AI User role assigned to Foundry project managed identity
  • Foundry resource identity
  • Bot Service principal
  • Scope matches the Foundry resource (not just the resource group)

Several users reported things working briefly, then failing again after redeployments.

6. Check Bot Service diagnostics (limited, but still useful)

If Application Insights is enabled on the Bot Service:

• Look for incoming activities from Teams/Web Chat
• Confirm the activity reaches the bot
• Note that logs often do not show why the response is dropped, but they help prove the request path is active

The lack of detailed diagnostics here is a known limitation today.

I Hope this helps. Do let me know if you have any further queries.

Thankyou!