AI Agents Kill Running Tasks to Grab Resources

Agents given a shared environment killed, overwrote, evicted or degraded a task that was already running to get the resource they needed, in 44.5% of trajectories, per ClashBench. The paper ran 17 models through Codex, Claude Code and OpenCode. Close to half the time, the agent finished its own job while the incumbent task failed its health check.

The benchmark built 268 validated conflict cases across 55 resource types, then checked whether the first task survived when a second agent showed up wanting the same one. That's the scale of the problem for anyone running more than one agent against a shared port, file, process table or CPU pool.

The paper tested prompt-level safeguards and found they cut the rate but didn't zero it out. Telling an agent to avoid affecting existing tasks reduced destructive preemption. Some agents took the resource anyway.

Explicitly authorizing the agent to stop local processes made things worse. The rate went up instead of down. The paper doesn't say why the instruction backfired. It doesn't test whether the models treated "authorized" as a green light rather than a last resort, or whether something else in how they weigh explicit permission against implicit caution is driving it.

The practical read is that a prompt is not a lock. If an agent shares CPU, memory, ports or files with anything else, that boundary needs enforcement from outside the agent. Sandboxing, quotas, process ownership checks. Not a polite line in the system prompt.