Altman and Amodei gave outsiders a seat and forgot the brake

On September 12, Anthropic gave an outside reviewer a desk: the same screen as the insiders, the right to report what they saw. Two days later, Sam Altman matched the pledge. Three rivals agreed over a weekend that outsiders should be able to look. Emad Mostaque named what the agreement left out: a reviewer who can see a departure and cannot stop the run is a witness. Maggie Appleton named a second gap, quieter but closer to your product: the agent starts work guessing what you meant, and you cannot see what it guessed until it has already acted on it.

These two gaps are the same gap at different scales. At the lab scale, a reviewer watches a run depart from the brief and has no way to intervene. At the product scale, a human hands work to an agent and the agent fills in the context it was never given, then proceeds. In both cases, the human's understanding and the agent's understanding diverge before the work begins, and the gap only becomes visible after the fact. Observability without interruption is a record of what went wrong, filed after the damage.

The design consequence for product teams is this: a seat and a screen are a starting condition, not a finished interaction. Appleton's proposal, a surface that shows what the agent understood before it acts, so the human can correct it, is the product-scale version of what Mostaque asked for at the lab scale. You need the correction to happen before the run, or you need a stop next to the seat. A reviewer who arrives at the end to assess what happened is doing recovery, not oversight.

The verb in Anthropic's spec is worth reading: verify, report, assess. All three happen after the run. The outside reviewer reads what was done and files a judgment. That is a review pattern, and review is valuable, but the interaction it describes is human-handoff: the agent finishes, the human judges. The interaction missing from every pledge made this weekend is interruption: the human sees the agent depart from the brief mid-run and redirects it before the departure compounds. Altman's caveat, pacing is not stopping, confirms the run continues while the reviewer watches.

Put the stop next to the seat. On your own product, before the agent starts a task with real reach, show the human what the agent understood: the goal it inferred, the constraints it assumed, the actions it plans to take first. Give the human a moment to correct the record. Then, during the run, make it possible to redirect without starting over. The outside reviewer at the lab and the operator at your product desk are doing the same job. Both need more than a screen.