What problem does it solve? Implementing multiple planned work items (beads) one at a time is slow, and naive multi-agent fan-out causes agents to trample each other's files, lose track of in-flight work, and skip review. This Skill coordinates a parallel swarm of implementation agents with file reservations, completion attestations, and a mandatory review gate. ## Core Features & Use Cases - NTM-first agent fan-out: Spawns implementation agents via ntm spawn and ntm --robot-send with marching-orders prompts, falling back to Agent() only when NTM is unavailable and the user approves. - Coordination and conflict prevention: Bootstraps Agent Mail sessions, team creation, and file reservations so parallel agents do not edit the same files. - Pre-completion quality gate: Requires each agent to run UBS checks, repo verify commands, self-review, and write a .pi-flywheel/completion/<bead-id>.json attestation before reporting done. - Wave-completion review gate: After all agents finish, routes the user through a consolidated review prompt (looks good, self review, fresh-eyes, or duel review) instead of ending the turn. - Use Case: After planning produces five ready beads, launch three parallel agents to implement them concurrently, monitor progress via the tending loop, and run the review gate once the wave completes. ## Quick Start Ask the agent to launch a swarm to implement all ready beads in parallel and choose how many agents should run.
