How It Appears in the Material Swarm review is a custom skill Ryan built for Notion AI workflows: a multi-agent review pass that vets generated code before a human looks at it. It appears as a step in his handoff prompt — after the agent fulfills requirements, it runs the swarm review skill for in-depth verification, then opens a PR and monitors CI. The principle: the human's job shifts to planning, orchestrating, and verifying, and much of the verifying can itself be delegated. [^7d2a9b32-b281-5828-9eb1-bf812129ae58@426] How the Swarm Works - Divide the change set into domains — e.g. frontend and backend. - Delegate per domain to two reviewer roles: a correctness reviewer hunting for errors, and a code maintenance reviewer looking for patterns to reuse, existing patterns that should be followed, and things that will fall apart at scale. - Run every pass on both GPT and Opus — two models per check. - A lead agent collects all findings and turns them into a report of what needs to be done. - Iterate until it's good — the loop repeats until findings are resolved. Why It Exists Big change-sets need more than bug-hunting — they need checks for maintainability. Rather than wait for someone to build the review skill he wanted, Ryan opened Codex and asked it to analyze other skills and create a new one, turning swarm review into his own thing. - Swarm review is a multi-agent verification skill that runs before human review. - It splits the change set by domain and uses correctness and maintenance reviewers, each run on two models. - A lead agent consolidates findings and iterates until they're resolved. - You can build your own skills with agents instead of waiting for them to exist.