New Foundry - Publishing an Agent to Web / Teams channel (No code) Results in 400 error

[alexanderdanson]

Hi,

I've attempted to publish a couple of agents on the new Foundry using a no code approach. In one case, I've connected a Knowledge Base using the built-in foundry tools (connected via Azure Search), and in the other case I've connected a custom MCP server. Both approaches work fine in the playground.

However, when I publish to Teams using the "one click publish", an Azure bot is created for me. When testing the web chat preview, and also after having deployed my bot as an app in Teams, I receive the same error:

Encountered error=Response status code does not indicate success: 400 (Bad Request).. If issue persists, please use following identifiers in any support request: ConversationId = XXXXX, activityId = XXXXX

I've added Application Insights, but the error seems to be happening before the agent runtime resulting in a lack of traces. I'm also writing from the Bot to my application insights resource. Log Analytics is also set up both for the Bot and for the agent, but doesn't provide any further details on the error.

I've attempted to solve by giving the Bot (Service Principal) a role of AI User / Developer on my Foundry resource in case this was an authentication issue but this didn't solve it.

Has anyone else experienced a similar issue and found a solution?

[BraydenJM]

Are you getting an 400 code when testing the web chat to the agent connected to the MCP server? I wanna make sure we're having the same issue. Using the new foundry portal and publishing an agent to an Azure bot im also getting 400 returns with no way (that I've found) of tracing this to see where something might be misconfigured

[alexanderdanson]

Yes exactly, I'm publishing it to an Azure bot (automatically created by Azure in the Publish to Teams workflow), and then I get the 400 code when testing the web chat. The same error occurs testing from Teams, so the issue seems to occur between the channel and the agent and doesn't seem dependent on the channel itself.

I have a separate agent which is only connected to a Foundry IQ Knowledge Base, no custom MCP, and I get the same issue when publishing to an Azure bot.

[smbakker]

This looks related to what I’m seeing.

When a published Azure AI Foundry Agent is invoked via the ActivityProtocol API, the agent runtime itself uses the ActivityProtocol execution path to send the reply activity. In this path, the callback activity contains an error message instead of a normal response.

The same agent configuration works correctly in inline chat / playground.

So the issue does not appear in inline chat, but only when the bot runtime responds via the ActivityProtocol path (in my case especially when Knowledge is enabled).

I’ve described the full reproduction and observations here for additional context:
https://github.com/orgs/azure-ai-foundry/discussions/279#discussion-9314840

[alexanderdanson]

Yes, that looks to be related.

One theory I had is if there's something required to automatically approve knowledge lookups and tool usage from channels and that this is causing the error (that there's no way to approve) - In the playground manual approval is required and this is one potential difference. However I can't find this setting on the agent in the new foundry to try it out. Do you have any ideas?

[smbakker]

Yes, that was one of our theories as well.

We’ve explicitly configured this in the agent YAML so that knowledge/tool usage does not require manual approval. After applying this change, the behavior remains the same: the agent works correctly in inline chat / playground (no approval needed), while the ActivityProtocol execution path still returns error messages.

So it doesn’t seem to be caused by missing approval configuration, but rather a difference in how the ActivityProtocol path handles knowledge/tool invocation compared to the playground.

[alexanderdanson]

Ok, thanks for letting me know - from what I can see the agent YAML is not editable when using the new Foundry UI, but if this doesn't solve the issue then I won't spend time looking into this. Let me know if you think of anything else, I shall do the same.

I'm wondering if there's something I've missed here relating to permissions for the published agent to access the tools:

If your agent includes tools that use Agent Identity for authentication, the newly created Agent Identity must have appropriate permissions
Navigate to each Azure resource your agent accesses and assign the required RBAC role to the new Agent Identity
The shared development identity permissions don't carry over—you must reconfigure permissions for the published agent's identity
After publishing, you can:
Open app to chat with your published agent application and easily share it with others in the UI (Note: sharing the application automatically grants them the Azure AI User role on the Agent Application resource)
Share the published endpoint with external consumers

[amynic]

Hi @alexanderdanson and @smbakker thanks for sharing your experience and feedback on the new foundry portal experience connecting to Web/Teams. Since New Foundry Portal came out of public preview we have made a repository for open feedback available here: https://github.com/microsoft-foundry/new-foundry-portal the team have asked if you can submit the an issue on this topic so they can assign and investigate

[alexanderdanson]

Thank-you, I have now submitted an issue here: microsoft-foundry/new-foundry-portal#71

[alexanderdanson]

@amynic For info, the auto-assign and automatic issue creation in Azure Devops failed when I created this issue due to some invalid token. If you need me to recreate the issue please let me know!

[smbakker]

@amynic We’ve added our own investigation details to the issue as well, including our findings around the ActivityProtocol behavior. Hopefully this helps with further analysis.

[amynic]

Thanks @alexanderdanson - that issues looks like it created successfully and the team are triaging the repo regularly. Thanks also @smbakker for adding extra details of your investigation this will help the team pinpoint any issues

Once we get feedback from the issue - lets update the thread with the outcome so others come across it if searching

[alexanderdanson]

@amynic Hi, do you have any updates on the progress of the issue?