Researchers have introduced Mid-Harness, a method that scales test-time compute for terminal agents by sampling and verifying candidate actions before one is forwarded for execution, while keeping the action generator and execution harness unchanged. The work addresses a key gap: a model's ability to generate a useful action does not guarantee that the action will execute reliably, and a poor command such as a wrong package install can alter the environment in ways that hinder later progress.
The paper's central finding is that verification quality determines how much agents benefit from additional action samples. With a TMAX-9B generator, more action sampling yielded little benefit under weak verification, whereas a capable verifier could exploit useful alternatives from the same generator.
On TerminalBench-Lite, a GPT-5.6 Sol verifier raised Pass@1 from 50.00% for the base agent to 68.03% with 8 sampled actions. When TMAX-9B itself served as the verifier, pairwise verification performed best among the evaluated verification mechanisms.
Distilling responses from the stronger verifier into TMAX-9B further improved Pass@1 while leaving the action generator unchanged. With TMAX-9B on TerminalBench-Lite, combining action and trajectory scaling reached higher success at lower estimated token cost than generating more trajectories alone.
The authors report that Mid-Harness also improves performance across additional models, benchmarks, and harnesses, identifying action scaling as a promising target for test-time compute scaling in terminal agents.