← sources

Our agent finished the task, but the caller never got an answer

source: reddit (reddit:AI_Agents)published 2026-10-02 14:10ingested 2026-10-03 02:35

retained text

On September 29, a teammate sent OmniNode a task. A routine redeploy restarted the services handling it one second later. The caller waited five minutes and timed out. The work finished about 23 seconds afterward, but its answer never got back.



We had marked the request as handled before returning the answer. During shutdown, we also closed the connection used for replies before sending the answers still waiting to go out. The caller had no way to tell whether the task had failed or was still running.



After the fix, an orderly restart returned a failure explaining that the runtime had shut down. A hard restart caused the request to be delivered again, and that attempt completed. We verified those two cases in the lab. That was enough to test the return path, not enough to call every retry safe.



If the first attempt had already changed a file or posted a comment, delivering the request again could repeat that action. We hadn't yet built the check for what the first attempt had already done.



For people running agents across restarts: how does the caller find out what happened when the worker finishes but the answer gets lost?

document fetched 2026-10-03 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1wvutnm/our_agent_finished_the_task_but_the_caller_never