Azure Function not working with Azure Agent Service

[justuswolff-audit]

As posted in this issue

I created an Azure function and connected it to my agent. When I run this code:

thread = agent_client.threads.create()
message = agent_client.messages.create(
    thread_id=thread.id,
    role="user",
    content="XYZ",
)

run = agent_client.runs.create_and_process(thread_id=thread.id, agent_id=agent.id)

I can see the request being queued in the input queue, processed and moved to the output queue, where it is eventually consumed from. However, the create_and_process function keeps on iterating in its internal loop and waiting for the function result. Citing from your sdk implementation:

# We need tool set only if we are executing local function. In case if
# the tool is azure_function we just need to wait when it will be finished.

Is there anything I am missing? Since the response is consumed, is there anything else I need to do?

[leestott]

Hi @justuswolff-audit

Potential Gotchas to Investigate further

Function Completion Signal Not Emitted: The create_and_process() loop internally waits for the function call to complete and it polls the run status. If the Azure Function completes but doesn't return a proper payload (or returns an empty or malformed tool response), the agent thread may stall waiting for a signal that never arrives.

Ensure your Azure Function posts a valid See https://learn.microsoft.com/en-us/azure/ai-services/openai/agents-reference#tool-call-response-format, especially with correct tool_call_id.

Tool Call Expectation vs Reality: Check that the tool name registered in the agent matches exactly what the function identifies as (especially if you used openai_tools.azure_function or a custom name).

If you haven't explicitly mapped the tool back to the function output with tool_choice or equivalent, the SDK might assume it still needs to run.

SDK Polling Timeout Configuration: create_and_process() has an internal loop with a default timeout, sometimes set to 10 minutes or more. If your Function runs slower, it's worth checking the agent run.status manually and inspecting run.steps for clues (agent_client.runs.retrieve(...)).

Alternatively, consider using the lower-level .create() followed by your own poll/check logic for finer control.

Missing or Misparsed Output Format: Did your Azure Function return a tool result in exactly the expected format? Even one level of JSON nesting off can be enough to derail the parser. Sometimes a {"output": "result string"} vs {"output": {"message": "result string"}} matters depending on how the agent tool expects it.

If you've already gone down these paths, feel free to share your function’s response payload structure.

[justuswolff-audit]

Solved, see issue.