Replies: 2 comments
-
|
Update We also investigated whether manual approval for knowledge/tool usage could be the cause. We explicitly configured the agent so that knowledge/tool usage does not require manual approval (via the agent YAML). After applying this configuration, the behavior remains unchanged: the agent works correctly in inline chat / playground, while the ActivityProtocol execution path still returns error messages. This further suggests the issue is not related to approval settings, but to differences in how ActivityProtocol handles knowledge/tool invocation compared to inline chat. |
Beta Was this translation helpful? Give feedback.
-
|
For visibility: our investigation and reproduction details have been added to We’ll follow up there going forward. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
TL;DR
When invoking a published Azure AI Foundry Agent via the ActivityProtocol API, callbacks return error messages instead of a normal response only when Knowledge (Azure AI Search) is enabled. The same agent works correctly in inline chat/playground. This appears specific to ActivityProtocol + Knowledge behavior.
Summary
This discussion concerns the Azure AI Foundry ActivityProtocol API (the protocol used to invoke published Foundry Agents programmatically).
The ActivityProtocol follows the Microsoft Agents / Bot Framework Activity schema, as documented in the public Activity Protocol specifications:
The issue described below focuses specifically on ActivityProtocol behavior in Azure AI Foundry, not on inline chat/playground usage.
I’m using Azure AI Foundry and observing different behavior between inline chat (playground) and the ActivityProtocol when an agent has Knowledge (Azure AI Search) configured.
202 Accepted.serviceUrl.This description is intentionally factual and avoids assumptions about internal implementation.
Environment
2025-11-15-preview*https://ai.azure.com)Agent configuration (important)
Reproduction steps
Create and publish an Azure AI Foundry agent.
Configure Knowledge (Azure AI Search) for the agent.
Verify the agent answers questions correctly in inline chat / playground.
Invoke the agent via ActivityProtocol with:
type: "message"conversation.idserviceUrlObserve:
202 AcceptedserviceUrlInspect the callback Activity payload.
Observed behavior
202 Acceptedwith empty body.type: "message"andfrom.id = "agent".textcontains an error message instead of a normal answer.Observed error variants
Variant 1 – Parsing-style error text
{ "type": "message", "from": { "id": "agent" }, "text": "'T' is an invalid start of a value. Path: $ | LineNumber: 0 | BytePositionInLine: 0." }Variant 2 – Generic HTTP 400 wrapper error
{ "type": "message", "from": { "id": "agent" }, "text": "Encountered error=Response status code does not indicate success: 400 (Bad Request).. If issue persists, please use following identifiers in any support request: ConversationId = <conversationId>, activityId = <activityId>" }Notes:
conversation.idequals the client-provided value.activityIdequals the sent message id (replyToId).Expected behavior
Question
Is this behavior expected or a known limitation of ActivityProtocol when Knowledge is enabled?
If so, what is the recommended protocol for chat-style integrations that require Knowledge?
If not, are there known issues or additional diagnostics that can be enabled?
Additional notes
Beta Was this translation helpful? Give feedback.
All reactions