Who reviews your agent's change: the author, the author in a fresh window, or someone else?

When your coding agent finishes a change and you ask it to look the change over, what is in the reviewer's window? If the answer is the whole session, the review is the author re-reading with its own notes in front of it: every file it opened, every dead end, and the moment it declared the change done. It knows why each line is there. A reviewer should not.

We wanted evidence that moving the review out of that window matters when the model stays the same. One small controlled study did exactly that: thirty artifacts with five planted errors each, one model, four review setups. The fresh window given nothing but the artifact caught the most, F1 28.6% against 24.6% for the full session and 40% of the critical errors against 29%, while a second pass inside the same session came last. Two agent vendors already build it that way: one CLI's guide puts it as a fresh context helping because the model "won't be biased toward code it just wrote", and the other vendor's team says one model does both jobs and was trained separately for each, with the reviewer running on the pull request.

Our own version reviews articles using a cold reviewer agent, handed that day's source material as ground truth. Each piece comes back with somewhere between one and twenty-three items we have to fix, usually four to six, and the hard ones were caught against the pasted sources.

If your agent reviews its own change, what does the reviewer get to see: the diff alone, the diff plus the task, or the whole transcript? And who has put two of those side by side on the same change?