# How should a NeMo agent handle a lost GPU job reply?

**URL:** <https://forums.developer.nvidia.com/t/how-should-a-nemo-agent-handle-a-lost-gpu-job-reply/384571>\
**Category:** NVIDIA NeMo\
**Tags:** cuda\
**Created:** [September 28, 2026, 12:23pm UTC](https://forums.developer.nvidia.com/t/how-should-a-nemo-agent-handle-a-lost-gpu-job-reply/384571 "2026-09-28T12:23:44Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![lukas.bower](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@lukas.bower](https://forums.developer.nvidia.com/u/lukas.bower)\
**Post date:** [September 28, 2026, 12:23pm UTC](https://forums.developer.nvidia.com/t/how-should-a-nemo-agent-handle-a-lost-gpu-job-reply/384571/1 "2026-09-28T12:23:45Z")

</div>

I’m integrating NeMo Agent Toolkit with an external GPU job executor. A tool reply can say the request was accepted, but the agent still needs to learn whether the host ran it and what result was observed.

In my Cohesix integration, those are separate states under one job ID. After a disconnect, the agent checks that job rather than submitting it again. How are others presenting an uncertain result to a NeMo agent: through separate MCP tool results, or one durable A2A task?