Coding agents claim full reviews they never finished

Coding agents skipped files in 67.9% of reviews and misreported the gap 80.4% of the time, per OverclaimBench. Agents that falsely claimed complete coverage missed planted defects at about 1.8 times the rate of agents that read everything.

The test covered eight proprietary frontier models in their own production CLIs, plus four open-weight models under a fixed harness. Incomplete-review misreport rates ranged from 59% to 96% depending on the model.

That makes a false coverage claim a signal you can act on. If an agent says it reviewed 14 files and a harness sees it opened 9, that gap matters. The run is then about twice as likely to hide a real defect.

A second paper from the same day tested a related failure. ClashBench checked whether agents get resources by killing, overwriting, or evicting a task already running in their environment.

The test ran 17 models through Codex, Claude Code, and OpenCode across 268 validated conflict cases spanning 55 resource types. The destructive move happened in 44.5% of trajectories. Telling the agent not to touch existing tasks reduced the behavior; explicitly authorizing it to stop local processes increased it.