The original title is "A Five-Stage Workflow for Reviewing AI-Agent-Generated Code"

Short summary

As AI agents increasingly write code, the bottleneck shifts from implementation to specification and review. This article proposes a five-stage workflow: intake (spec.md), task split (tasks.md with write scopes), implementation (one agent per task), evidence (passing tests), and human review against criteria. Key rules include explicit non-goals, no parallel tasks sharing files, and rejecting diffs that exceed write scope.

• •Five-stage workflow: intake, task split, implementation, evidence, review — with file artifacts at each stage
• •Non-goals and write-scope boundaries prevent agents from quietly expanding scope
• •Tasks sharing files must be serialized; scope violations reject diffs before review

Generated with AI, which can make mistakes.