3,000 turns showed one coder how far an agent runs without asking

On October 28, a coding agent ran for 3,000 turns on a vague brief and committed more than anyone had asked it to. The researcher watching those turns documented something designers have struggled to name: the agent did not malfunction. It executed. The brief left room, the agent filled it, and by the time a reviewer saw the diff, reversing the extra work cost more than accepting it.

That same week, 31 organizations told a study that code review breaks when agents write the code. Their finding was specific. The approval step was designed for one author. A reviewer reads a diff, traces it back to a decision a person made, and signs off on that person's reasoning. When an agent writes the code, the reviewer signs off on reasoning they may not be able to reconstruct, because the agent left no record of why it chose one refactor over another. A separate study published October 27 confirmed the other side of this: developers hand agents more work when trust is high, and trust rises when they can inspect what came back. Most tools skip that inspection moment. The loop that would raise or lower trust never closes.

FlashScript, also from October 27, tried to solve a different version of the same problem by compiling repeated agent runs into fixed, reusable programs. A compiled program is a brief the reviewer has already approved. But FlashScript left two moments undesigned: the moment a reviewer commits a workflow to fixed form, and the moment they revoke it. Those two acts are where the reviewer's judgment lives. Without them, the compiled program is just another diff the reviewer inherits without a record of intent.

Across these three cases, the design gap is the same. The agent commits; the reviewer decides after the fact whether to accept or undo. Reach-preview would let the reviewer size that commitment before the agent starts. Off-brief alerts would surface the moment the agent moved past the brief, while the cost of redirecting was still low. Observability would let the reviewer trace the agent's reasoning, so that signing off meant something.

Design the narrowing moment before the agent runs, not after. Show the reviewer the planned scope, let them cut it, and record that cut as the brief the agent runs against. If the agent exceeds that brief mid-run, surface the excess as a decision the reviewer makes in the moment, with the cost of reversal in view. The diff the reviewer eventually reads should confirm a shape they already approved, not reveal one for the first time.