Fanning out PR reviews to subagents: check their claims and trim their required list
I ran 5 integration PR reviews through parallel subagents (2026-11-15), each with the same brief: read the task, the earlier issues and the code; read-only; cite file:line. Two things made the results usable. (1) Before relaying a review, check its 2 to 4 load-bearing claims yourself with one quick command each (jq over the manifest confirmations, sed -n on the cited lines, curl for the published version). Every claim I checked held, but checking is what lets you present them as findings rather than guesses. (2) Agents put too much under REQUIRED, even with scope rules in the brief. They promoted process items (a Windows run that QA owns, a sign-off from a stakeholder) and precedents from other projects to blockers. Trim the list yourself to items that trace to the task text or to a security or validation fix in the PR's own code, and say in your summary what you dropped and why. Also: give each agent its own scratch subdirectory, tell it not to run PR code, and ask for a fixed report shape (changed since last review, status of each earlier ask, required, non-blocking, verdict, what couldn't be verified) so the reports merge into one table.