iterative-fleet

What problem does it solve?

iterative-fleet helps teams run complex, multi-round work with multiple parallel agent workers while enforcing a quality gate so the work only advances when an explicit reviewer verdict approves it.

Core Features & Use Cases

• Reviewer-gated iteration loop: Workers run in cycles and a dedicated reviewer reads worker logs to write a verdict that determines whether to continue, iterate again, or pause for escalation.
• DAG ordering per iteration: Uses depends_on to run workers in topological layers for correct sequencing (e.g., researcher → builder → reviewer) across every iteration.
• Multi-provider worker support: Designed for headless claude -p, codex exec, or pi -p providers, with configurable model settings per fleet/worker.
• Operator-controlled safety: The orchestrator never auto-kills or auto-restarts workers; it only reads verdicts and worker completion signals, so operators remain in control of intervention.

Quick Start

Launch an iterative reviewer-gated fleet by creating a fleet.json with exactly one type: "reviewer" worker and then running bash ${AGENTS_SKILLS_DIR}/scripts/launch.sh $FLEET_ROOT followed by bash ${AGENTS_SKILLS_DIR}/scripts/status.sh $FLEET_ROOT.