Peer review organized one reviewer per artifact is a checkpoint; organized as a network — rotating reviewers, blind passes, calibration against seeded defects — it becomes a quality system with measurable catch rates. Agent code review is where this matters most, fastest.

The design borrows deliberately from how human review works at scale. Reviewer rotation prevents the single-reviewer blind spot: the same reviewer-model reviewing the same maker-model develops the same blind spots, politely approving the same classes of bug.Blind review — the reviewer sees the artifact and standards, not the author or its reasoning — removes the anchoring that makes reviewers lenient toward work they watched being made. Calibration sets — artifacts with known, seeded defects injected at known severities — convert "the reviewer seems good" into "this reviewer catches 7 of 10 critical classes, misses concurrency, over-flags style." No serious review system ships without the last one.

The pipeline that emerges: maker produces code with tests; first-pass reviewers run in parallel under different lenses — correctness, security, performance; findings are deduplicated and severity-classified; the maker responds to each finding with a fix or a reasoned rejection; contested findings escalate to an arbiter pass. The metric that rules everything is catch rate per defect class over time — reviewed against the seeded ground truth, not against vibes.

Lessons That Transfer Directly

• Review the diff plus its blast radius: the changed lines and the things they touch — interface consumers, config, tests that assert the old behavior.
• Findings are typed: blocking correctness, security, and data-loss findings differ from preference notes; conflating them trains makers to ignore reviewers.
• Track reviewer agreement: when two reviewers split on a finding, that split is a calibration datum — route it to the arbiter and learn which lens was right.
• Never let the maker merge: the review network's authority ends at a promotion gate the maker cannot reach — the same fence, drawn through the review graph.

Peer-review networks are also the cheapest place to start with multi-agent systems: two agents, a checklist, and a seeded-defect harness produce measurable quality movement in a week — no exotic orchestration required.