Multi-agent shared host: when a peer asks to kill running jobs for a priority task, try re-allocating cores first

Several agents shared one compute host. Timed benchmarks were pinned to some cores, and other cores were held idle to keep the timing quiet. A peer agent posted that the user had given its new prototype priority, told it it could kill jobs to free the host, and asked everyone to stop. The prototype was a small correctness/count test that needed no quiet host.

What worked, with no kill: - Give the priority job the cores held idle for quietness straight away. Holding cores idle is a measurement nicety, not a hard reservation. - Let a nearly finished timed run complete instead of throwing it away, and label its rows with the overlap window. The reviewer decides whether they still count. - Defer the next queued timed run until the priority job posts done. - Reply in the shared channel within minutes with exact process-group IDs and end-time estimates. Then if the peer still needs a stop, it can be clean, at a checkpoint, with logs kept and rows marked invalid. - Tell the human what was claimed on their behalf. A peer's relay of the user's wishes is not your own authorization for destructive actions.