An evolve agent rewrote a coding agent's harness and beat the hand-built Codex harness

AHE is a closed loop in which one agent edits the harness of another while the base model stays fixed. The harness is decomposed into seven component types held as plain files: system prompt, tool descriptions, tool implementations, middleware, skills, sub-agent configs, and long-term memory. Rollout trajectories are distilled into a layered evidence corpus, and every edit ships with a manifest naming the failure evidence it targets, the tasks it predicts it will fix, and the tasks it puts at risk. The next round grades that prediction, and edits that do not cash out are reverted at file granularity. Starting from a bash-only seed, with GPT-5.4 driving all three roles, ten iterations lifted pass@1, the single-attempt solve rate, on Terminal-Bench 2 from 69.7 percent to 77.0 percent, past the human-designed Codex harness at 71.9 percent and two self-evolving baselines, ACE and Training-Free GRPO. The frozen harness then transferred without further evolution: highest aggregate success on SWE-bench-verified at 12 percent fewer tokens than the seed, and gains of 5.1 to 10.1 points across three other model families. Ablations put the gain in tools, middleware, and long-term memory; the evolved system prompt inserted alone scored below the seed. The loop's weak side is regressions. Fix predictions landed roughly five times above a random baseline, regression predictions only about two, and the authors name "regression foresight" as the clearest direction for future loops while calling the system a controlled research prototype. Code and the full prompts are public.

Why it matters: Practitioners still tune harnesses by hand. This paper offers an auditable way to delegate that work: require each edit to state a testable prediction and roll it back when the prediction fails. Its ablation also points to where tuning time paid off. The transferable value sat in tools, middleware, and memory rather than prompt prose.