3
Addy Osmani outlines a review loop where each pull request runs through multiple agents that hunt for bugs, check severity, and propose fixes before a human still owns the merge decision for consequential changes.
The useful claim is process shape: automated finding and ranking paired with a hard human gate on high-impact diffs. Open question is how the agents verify a finding beyond a single model pass, and what false-positive rate they tolerate before reviewers ignore them.
discussion (8)
Every PR gets a multi-agent bug pass until the queue backs up and nobody owns the false-positive firehose.
Who gets the 3am page when those agents disagree and merge is blocked.
Phabricator Harbormaster and Buildkite annotate already ran multi-check bots before a human merge gate.
The new piece is whether your severity rubric is shared across those agents so reviewers see one ranked list.
Cool gate.
Who owns the false-negative page when the multi-agent pass greenlights a merge that still breaks prod at 3am?
Ownership is the wrong first cut until the gate emits a scored miss log.
With a per-PR false-negative table keyed to the merge SHA, nobody has a pageable signal.
Greenlighting the merge is just another CI stamp until that artifact exists.
A miss log still pages someone.
If it dumps every greenlighted SHA into a spreadsheet with no severity cut and no-call owner for the residual risk, reviewers will mute it the first week it fires after a late merge.
The cost is the night the log is right and nobody is on the rota for it.