## Definition The **Quorum Review Pattern** is a multi-agent composition in which N independent reviewer agents each produce a vote on a diff, and the diff proceeds only when a threshold number of votes pass (e.g., 2-of-3). It is designed to catch single-agent bias that a 1:1 [[Builder-Critic Pattern]] cannot detect: one reviewer may share the builder's blind spot, but three diverse reviewers are unlikely to share the same one. ## Topology ``` [Builder] → diff → [Reviewer-A (Sonnet)] → vote: PASS / FAIL → [Reviewer-B (Opus, ultrathink)] → vote: PASS / FAIL → [Security Auditor (Opus)] → vote: PASS / FAIL If votes_pass >= 2 → proceed to [Human] If votes_pass < 2 → return to [Builder] with merged feedback from all failing reviewers ``` The author's formulation from Module 7: *"run N independent reviewers and require consensus."* ## Distinction from Builder-Critic [[Builder-Critic Pattern]] is a 1:1 topology: one builder, one critic, with revision loops capped at 3. The Quorum Review Pattern is a 1:N topology: one builder, N independent reviewers voting in parallel. The two patterns differ in: - **Purpose.** Builder-Critic catches implementation drift from the spec. Quorum catches single-reviewer bias on high-stakes semantics. - **Cost.** Quorum costs N× the reviewer tokens per cycle, making it inappropriate for routine changes. - **Diversity.** Quorum gains power from using heterogeneous reviewers (different models, different tool surfaces, different specializations). ## When to Use Overkill for most changes. The author's trigger list from Module 7: - Schema migrations. - Authentication and authorization logic. - Billing logic. - Any change you would page someone for at 3 am. A useful heuristic: if a [[Builder-Critic Pattern]] PASS would still make you anxious, reach for Quorum. ## Reviewer Diversity Axes To avoid the same blind spot appearing in all reviewers, vary along at least two axes: 1. **Model family or capability.** Sonnet + Opus. 2. **Reasoning mode.** Standard + `ultrathink`. 3. **Specialization.** General reviewer + Security Auditor + domain expert. ## Feedback Aggregation When the quorum fails, merge the failing reviewers' feedback before returning it to the builder. Duplicate concerns should be deduplicated; conflicting concerns should both be surfaced. The builder addresses the union of all critiques, not just one reviewer's list. ## Related - [[Builder-Critic Pattern]] - [[Verifier Independence]] - [[Specialized Agent]] - [[Adversarial Agent]] - [[Sequential Pipeline]] ## Sources - [[Modern AI Software Engineering - The Orchestrators Playbook]]