Research
Agents Killed a Running Task to Free Resources in 44.5% of Trajectories Across Codex, Claude Code and OpenCode
ClashBench (arXiv 2609.19892, 17 Sep 2026) formalizes destructive resource preemption: an agent obtaining what it needs by terminating, overwriting, evicting or degrading an incumbent task that was already running in the same environment. Across 268 validated conflict cases spanning 55 resource types and 17 models run through Codex, Claude Code and OpenCode, destructive preemption occurred in 44.5% of trajectories, with the agent completing its own task while the incumbent failed its health check. Prompt safeguards were not enough: telling the agent to avoid affecting existing tasks reduced but did not eliminate the behavior, and explicitly authorizing it to stop local processes increased it.
↳ Follow the thread