A middle manager for your AI coding agents

What it does

Boss Skill is a TypeScript CLI runtime that plugs into Claude Code, Codex, OpenClaw, and similar agents to run a full software-development assembly line. You give it a sentence like “build a todo app,” and it spins up nine role-specific agents—PM, Architect, Frontend, Backend, QA, DevOps, and others—moving them through four rigid stages from requirement clarification to deployment. The runtime tracks state in an event-sourced journal, enforces quality gates that block release on lint errors, test coverage, or Lighthouse scores, and can resume interrupted runs by fingerprinting artifacts to decide what to reuse or rerun.

The interesting bit

The clever part is the “conversation layer” between agents: instead of silently overwriting files, an agent can challenge, propose, or escalate to another, and every resolution must land as a single-owner todo before work continues. It is essentially a workflow engine treating LLM agents like microservices that must check in with a central state machine.

Key highlights

• Plays nicely with multiple host agents—Claude Code, Codex, Antigravity, Hermes—installing itself as a skill or plugin rather than replacing them.
• Four-stage pipeline with non-bypassable gates (TypeScript/lint, 70% coverage / zero P0 bugs, Lighthouse / P99 latency).
• Event-sourced runtime: state changes append to .meta/events.jsonl, and the CLI can replay history or resume a past run by hashing workflow nodes to decide reuse, skip, or rerun.
• Human-in-the-loop levels range from fully automatic to interactive, with a “lightweight” mode that trims the cast to four core roles.
• Hooks guard the full agent lifecycle—session start, file writes, bash commands, child-agent dispatch—controlled by environment variables.

Verdict

Worth a look if you are already using Claude Code or Codex and want structured guardrails rather than freestyle coding. Skip it if you prefer tight, manual control over every commit and architectural decision.