Review Gate Patterns

A review gate is a checkpoint that output must pass before advancing to the next stage. It is the production line's quality control station. Every deliverable stops at the gate, gets inspected against defined criteria, and either passes through or gets sent back for rework. Without gates, the pipeline is a conveyor belt with no quality control — whatever the first agent produces flows straight to the final deliverable.

Gates come in three varieties. Automated gates use scoring thresholds: "All dimension scores must be 7 or above to pass." These run without human involvement and handle the majority of quality checks. Conditional gates apply different thresholds based on context: "Internal documents pass at 6+. Client-facing documents pass at 8+. Board presentations pass at 9+." The stakes determine the bar. Human-in-the-loop gates require explicit human approval before proceeding: "Legal review of contract language requires human sign-off regardless of automated scores." Some decisions should not be delegated to agents.

Gate placement is a design decision. Too few gates and errors propagate. Too many gates and the pipeline stalls. The principle I follow: place a gate after every stage where the output format changes. Research produces structured data — gate. Drafting produces prose from that data — gate. Review produces feedback — no gate needed, feedback goes directly to the revision stage. Formatting produces the final deliverable — gate.

Format changes are natural quality checkpoints because they are the points where information can be lost or distorted. When structured data becomes prose, did the key findings survive the transformation? When a draft becomes a formatted PDF, did the formatting preserve the content hierarchy? These are the junctions where quality review has the highest impact.