Mid-Harness: verify the action before the environment commits
Sep 30 Kang et al.: sample N candidate terminal actions and verify at the model–harness boundary—GPT-5.6 Sol verifier lifts TMAX-9B Pass@1 50→68% on TerminalBench-Lite; pairwise + distillation recover part of the gap without retraining the generator.
Source: Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents
Minki Kang and collaborators target a failure mode that overnight loops feel acutely: a single bad command mutates the environment (wrong package, wrong edit) long before the trajectory score turns red. Trajectory-level Best-of-N and sequential refine help, but they pay for whole re-runs. Mid-Harness inserts compute at the model–harness boundary: at each step, sample $N$ candidate actions from the same history, verify them, and forward only the chosen one—generator weights and harness loop unchanged.
The coverage claim is strong. With TMAX-9B fixed on TerminalBench-Lite, a frontier listwise verifier (GPT-5.6 Sol) raises Pass@1 from 50.00% → 68.03% at $N=8$, evidence that useful alternatives already sit in the sample and weak verification leaves them on the floor. Self-verification is harder: more candidates barely help under weak listwise scoring; pairwise comparison is the best of the evaluated mechanisms (54.76% Pass@1), and distilling the frontier verifier’s pairwise judgments into TMAX-9B lifts that to 57.14% without touching the generator. Combining Mid-Harness with Best-of-$T$ or Sequential Refine beats spending the same budget on more full trajectories alone—and the pattern transfers across model sizes, Terminus-2 / Vanillux2 harnesses, Terminal-Bench 2.1, SWE-bench Verified Mini, and FeatureBench-Mini.
Why it matters for code review
Continuous verification under volume is not only CI after the PR—it is the pre-execution gate inside the turn. Mid-Harness makes that gate reviewable: which mechanism (listwise / pointwise / pairwise), what false-accept rate on command semantics and execution feasibility, and whether action scaling is cheaper than another overnight re-run. Productive complication: without a capable verifier, sampling width is mostly theater. Pair with HiSentinel’s Allow/Redirect/Hard-Pause pre-execution policy and AgentRelay’s evidence-gated done.