FailFast monitor saves 20.4% of agent tokens

A 0.6B-parameter FailFast monitor saved 14.6% to 20.4% of execution tokens across coding-agent policies on SWE-bench Verified at a 5% false-positive target. The RestartSmart study also reports that, at a more aggressive 25% false-positive target, restarting Qwen3.6-27B with the interrupted repository diff available raised resolution from 66.6% to 71.8%, versus 66.8% for a cold restart. The finding sharpens the case for treating stopping rules as part of the agent harness: log prefix telemetry and preserve useful edits before paying for another full trajectory, but calibrate the monitor against the cost of killing a run that would have succeeded.