Gas Town is a multi-agent orchestration system that coordinates multiple AI coding agents working on different tasks within a persistent workspace.
The core problem Gas Town addresses is the loss of context and work state when AI agents restart or hand off tasks to one another. Traditional agent workflows struggle with coordination across multiple agents and lack reliable persistence. Gas Town solves this by storing work state in git-backed hooks and providing a structured workspace architecture. The system introduces a hierarchical organization: a Mayor (primary Claude Code coordinator) oversees a Town (workspace directory) containing Rigs (project containers with git repositories). Within this structure, Polecats are worker agents with persistent identity and work history despite ephemeral sessions. Work is tracked through Beads, a git-backed issue system that stores structured work state, and organized into Convoys that bundle multiple beads for assignment. The Refinery component handles merge queue processing and verification gates when agents complete work, while a three-tier watchdog system (Witness, Deacon, and Dogs) monitors agent health and triggers recovery when needed.
Gas Town suits teams running 20 to 30 coordinated AI agents on complex projects where context loss and manual coordination become bottlenecks. It is particularly valuable for workflows requiring reliable handoffs between agents and persistent tracking across restarts. The system scales beyond what manual agent management can handle, making it appropriate for organizations treating AI agents as a core part of their development infrastructure rather than occasional tools.
The project shows active development with ongoing refinement of its core concepts and architecture. Work is organized through a structured tracking system using Beads and Convoys, indicating the maintainers are using their own tool for development. The system includes sophisticated features like Molecules for workflow templating with multiple execution modes and Escalation for handling failures, suggesting iterative expansion of capabilities based on real-world usage patterns.
