0.6B monitor kills failing coding agents, lifts fixes to 71.8%

A 0.6-billion-parameter monitor watches a coding agent's trajectory and flags, before the run ends, that it's headed for failure, per a new arXiv paper describing a method called FailFast-RestartSmart.

Killing a doomed run early cuts execution tokens by 14.6 to 20.4%. But the bigger number comes after the kill. Restarting with the abandoned diff carried forward, as an optional starting point, lifts the resolve rate to 71.8% on SWE-bench Verified.

The monitor only sees the trajectory itself, no policy logits, no hidden states. That means it works as an outside wrapper, not a change to the underlying model. It was trained only on traces from Qwen3.6-27B, then transferred without retraining to three other policies, including one closed-API model, per the paper.

A cold restart with no diff carried forward reaches 66.8% on the same benchmark, five points below the full method's 71.8%. That gap is the entire case for the 'RestartSmart' half of the name: keeping the diff, not stopping the run, produces the accuracy gain.

For agentic coding, that means recovery design is doing more work than failure detection. A monitor that kills bad runs but restarts from a blank slate would capture the token savings and lose most of the accuracy gain. Anyone building agent harnesses should treat an interrupted diff as reusable state, not scrap, and watch whether the cross-policy transfer holds up outside this benchmark.