Researchers present Mid-Harness, a method that allocates test-time compute at the boundary between a language model and its execution harness for terminal-based AI agents. Rather than changing the underlying model or harness, the system samples multiple candidate actions and uses a verifier to select one before execution, addressing cases where a single poor command can derail an agent’s progress. On TerminalBench-Lite, pairing a TMAX-9B action generator with a GPT-5.6 Sol verifier raised Pass@1 from 50.00% for the base agent to 68.03% using eight sampled actions. The authors also show that distilling the stronger verifier’s judgments back into the smaller model improves performance while leaving the action generator unchanged, and that combining action- and trajectory-level scaling outperforms generating more full trajectories at a given token budget.
