Ship a 0.6B failure monitor next to your SWE agent: it cuts 14.6–20.4% of execution tokens and lifts resolve rate 66.6%→71.8% by restarting instead of grinding

arXiv 2608.03222·high signalXBlueskyLinkedInCopy link

FailFast-RestartSmart trains a 0.6B monitor on observable trajectory prefixes only — no policy logits or hidden states — to predict that a repo-level agent run is going to fail, then launches a fresh same-policy rollout with the interrupted diff offered as an optional overlay the agent may inspect, apply, or discard. Trained solely on Qwen3.6-27B traces it transfers to three other policies including a closed-API model, saving 14.6–20.4% of execution tokens at a 5% false-positive target, beating a per-step AgentStop adaptation (12.5%). The actionable part for builders: cold restart alone only reaches 66.8% on SWE-bench Verified, while carrying the diff forward reaches 71.8% — the recovery, not the kill, is where the gain lives.