Argus.
Three-agent orchestration platform. Claude plans, Gemini builds, Codex audits. Coordinated by Hermes.
Built a three-agent orchestration platform that turns raw LLM outputs into production-ready code by enforcing a planning step, a building step, and an independent audit, with a human review loop when the audit isn't clean. Claude writes a `<slug>-Plan.md` with architecture and acceptance criteria; an opt-in plan-review gate pauses the pipeline so the human can approve or send feedback before Gemini implements. Gemini then logs each iteration; Codex audits across five categories (Security, Correctness, Resource, Interaction, Maintainability) and grades A/B/C/F. Non-A grades route back to Gemini with the audit feedback and a frozen plan; an opt-in auto-approve loop runs build → audit → revise unattended up to a configurable cap.
The orchestration engine, Hermes, runs XState v5 state machines coordinated over NATS pub/sub with SQLite persistence and a chokidar file watcher that converts file writes into state transitions. React 19 + Vite + Tailwind v4 cockpit ships a live agent-color progress strip, real-time output stream, paginated logs with iteration drill-down, and a Warzone discussion mode where all three agents debate an idea before any code is written.