Azure AI SDK + Streamlit: Event loop closed error in threaded async wrapper

[PeaceES]

Hi there,
(Using Azure AI Projects SDK (v1.0.0b12) + Azure AI Agents SDK (v1.0.2) in Streamlit app)

My agent works perfectly in terminal but fails in Streamlit with "event loop is closed" error.

Is there a recommended pattern for using Azure AI SDK in Streamlit or other threaded environments? The issue seems to be HTTP client cleanup timing when the event loop closes

Attempted solutions that didn't work:

• different async wrapper patterns (asyncio.create_task, event loop management)
• streamlit's @st.cache_resource for persistent clients
• manual HTTP session cleanup
• various ThreadPoolExecutor configurations

Please see screenshots below and many thanks in advance!

[leestott]

Hi @PeaceES

So for Azure AI Agents with Streamlit’s synchronous runtime. The event loop is closed error typically stems from Streamlit reusing or prematurely closing the loop across reruns or threaded callbacks.

Here's a distilled pattern that’s been shown to work more reliably:

Recommended Pattern: Isolated Event Loop Execution

Use a dedicated event loop per async call, and catch loop closure errors gracefully.

import asyncio

def run_async_task(coro, *args):
    try:
        loop = asyncio.new_event_loop()
        asyncio.set_event_loop(loop)
        return loop.run_until_complete(coro(*args))
    except RuntimeError as e:
        # Retry with fresh loop if closed
        loop = asyncio.new_event_loop()
        asyncio.set_event_loop(loop)
        return loop.run_until_complete(coro(*args))
    finally:
        loop.close()

This avoids reusing a stale loop and sidesteps cleanup timing issues from httpx or aiohttp clients used under the hood by Azure SDKs.

Additional Tips

• Avoid asyncio.run() inside Streamlit—it closes the loop after execution.
• Don’t cache async clients with @st.cache_resource unless you wrap them in sync-safe accessors.
• Use run_async_task() for every async call, especially when invoking agents or threads.
• Close clients explicitly if they expose .close() or .aclose() methods, but only after loop completion.

Here is example code for Agent Calling in Streamlit

Please try the following

def query_agent(agent_client, thread_id, message):
    async def run():
        await agent_client.create_message(thread_id, role="user", content=message)
        return await agent_client.create_run(thread_id, agent_id)

    return run_async_task(run)

Some additional reading
https://learn.microsoft.com/en-us/answers/questions/1465989/how-to-get-streaming-response-for-chat-completion
https://stackoverflow.com/questions/45600579/asyncio-event-loop-is-closed-when-getting-loop

[amynic]

@PeaceES if @leestott answer has helped you please 'Mark as answer'