Intermittent SSL Timeout / Incomplete Stream Error When Supervisor Agent Orchestrates Multiple Agents Issue Description

[Keerthanamurugesan2001]

I am facing intermittent SSL timeout / connection termination errors when using a Supervisor Agent that orchestrates multiple child agents (3 agents) using Agent Framework + AG-UI.

The issue does not happen consistently:
Sometimes the workflow runs successfully
Sometimes it fails mid-execution with SSL / streaming errors

What Works
Individual agents run fine when executed alone
Supervisor agent works sometimes

Has anyone faced SSL timeout or incomplete stream errors and how to fix it.

[moonbox3]

Hi @Keerthanamurugesan2001,

To help diagnose this, we need more details:

1. Which SDK? Python or .NET?

2. Agent Framework version? (for example, with Python, pip show agent-framework or check your pyproject.toml)

3. AG-UI version?

4. LLM provider? Azure OpenAI, OpenAI, Anthropic, etc.?

5. Deployment environment? Local dev, Docker, Azure App Service, Kubernetes, etc.?

6. Network setup? Behind a corporate proxy, VPN, or firewall?

7. The actual error message/stack trace: please share the full traceback.

8. Minimal repro code — a stripped-down version of your supervisor setup that triggers the error.

9. Does the issue correlate with:

   • Longer-running agent tasks?
   • Higher concurrency (all 3 agents active simultaneously)?
   • Specific network load?

[Keerthanamurugesan2001]

Thank you for your response here is the details you asked:

1. Which SDK - I am using Python,
2. Agent Framework version - 1.0.0b251120
3. AG-UI version - Latest version 0.0.42
4. LLM provider - Azure OpenAI
5. Deployment environment: Local dev
6. Network setup? Corporate proxy (VDI)
7. Full traceback:

INFO:agent_framework_ag_ui._orchestrators:[FINALIZE] Checking for unclosed message. current_message_id=77f31077-a24f-4be7-a4d0-5b2f2d8ca70a
INFO:agent_framework_ag_ui._orchestrators:[FINALIZE] Emitting TextMessageEndEvent for message_id=77f31077-a24f-4be7-a4d0-5b2f2d8ca70a  
INFO:agent_framework_ag_ui._orchestrators:[FINALIZE] Emitting MessagesSnapshotEvent with 11 messages (text content length: 332)        
INFO:agent_framework_ag_ui._orchestrators:[FINALIZE] Emitting RUN_FINISHED event
INFO:agent_framework_ag_ui._orchestrators:Completed agent run for thread_id=dd0acd91-7dc2-4a71-90b5-2c2e437ae57e, run_id=625d68ab-e5b4-411d-a914-8641cc7493d9
INFO:agent_framework_ag_ui._endpoint:[/supervisor] Completed streaming 142 events
ERROR:asyncio:Future exception was never retrieved
future: <Future finished exception=ClientConnectionError('Connection lost: SSL shutdown timed out')>
TimeoutError: SSL shutdown timed out

8. Minimal repro / context:

I don’t have a small standalone repro yet. The issue occurs in my full setup:
1 supervisor agent
3 worker agents:

• Extraction agent
• RAG agent
• Checklist agent

All three agents can be active concurrently, each running tool calls that process large payloads (document extraction, retrieval, and validation).

9. Observed correlation

Yes, the issue correlates with:
Higher concurrency — when all 3 agents are active simultaneously
Longer-running tasks — especially heavy extraction and RAG steps
⚠️ Network conditions — more frequent failures when running behind a corporate proxy / Citrix environment

[moonbox3]

Thanks for the added details. It's going to be difficult for me to exactly pinpoint the issue as I can't reproduce it given you have a corp proxy / Citrix env. But some thought are:

I believe the error ClientConnectionError('Connection lost: SSL shutdown timed out') is happening during connection cleanup, not during the actual request. The "Future exception was never retrieved" indicates this is a background task failure during graceful SSL termination.

My thinking is that your corp proxy (Citrix) is terminating idle connections aggressively. When agents do long processing between LLM calls, connections go idle, proxy kills them, and when the client tries to gracefully close SSL, the connection is already gone.

One thing worth trying, configure the Azure OpenAI client with a shorter keep alive and explicit timeouts:

import httpx
from agent_framework.azure import AzureOpenAIChatClient
from openai import AsyncAzureOpenAI

client = httpx.AsyncClient(
    timeout=httpx.Timeout(60.0, connect=10.0),
    limits=httpx.Limits(
        max_keepalive_connections=5,
        max_connections=10,
        keepalive_expiry=30.0,  # shorter than proxy timeout ideally
    ),
)

chat_client = AzureOpenAIChatClient(
    # ... your config
    async_client=client,  # the custom AsyncAzureOpenAI from above
)

You may also experiment with toggling http2=False in the httpx.AsyncClient(http2=False) config, as disabling HTTP/2 is not handled as well as HTTP/1.1 when using a proxy.

As I mention above, the "Future exception was never retrieved" is most likely a cleanup issue: the actual agent run likely completed. After experimenting with the custom httpx client, and you continue to see errors, you could try suppressing these during shutdown, and wrap your agent run in proper cleanup:

import asyncio

async def run_with_cleanup():
    try:
        await your_agent_run()
    finally:
        # Give pending SSL shutdowns time to complete or fail
        await asyncio.sleep(0.5)

Through all of this, adjusting the keepalive_expiry should hopefully help. You may want to experiment with reducing the amount if you continue to hit errors. Let us know if any of this helps, or you continue to experience issues.