# Gas Town Gas Town is a multi-agent orchestration framework created by [[Steve Yegge]], written in Go (~189k lines). Published on January 1st, 2026, it enables developers to manage colonies of 20-30 parallel [[AI Agents]] (primarily [[Claude Code]] instances) through a structured hierarchy of roles, workflows, and persistent state management. Yegge describes it as "Kubernetes mated with Temporal" but focused on agent orchestration. The system uses a Mad Max-themed naming convention throughout. ## Architecture Gas Town operates with a two-level organization: - **Town** (`~gt` directory): Headquarters managing configuration and all projects - **Rigs**: Individual project repositories under the Town's control, each running specialized agents ## Agent Roles ### Town-Level (Cross-Project) - **Mayor**: Chief dispatcher orchestrating major workflows and handling user requests - **Deacon**: System health daemon running maintenance patrols - **Dogs**: Helper agents assisting the Deacon ### Rig-Level (Project-Specific) - **Crew**: Named, persistent agents handling design and review work - **Polecats**: Ephemeral workers spun up for specific tasks then terminated - **Refinery**: Manages merge queues and handles conflicting merges - **Witness**: Supervises Polecats and helps unstick blocked work ## The MEOW Stack (Workflow Layers) Hierarchical work abstraction enabling crash recovery: 1. **[[Beads]]**: Atomic work units (JSONL-based issues tracked in Git) 2. **Epics**: Hierarchical collections of Beads 3. **Molecules**: Instantiated workflow graphs with dependencies and gates 4. **Protomolecules**: Reusable workflow templates 5. **Formulas**: High-level [[Tom's Obvious Minimal Language (TOML)]] definitions "cooked" into protomolecules ## Key Concepts - **GUPP** (Gas Town Universal Propulsion Principle): If there is work on your hook, you MUST run it. This ensures persistent scheduling; agents check hooks and resume workflows after crashes - **Hooks**: Bead-backed queues where molecules attach; work persists via Git, enabling resumption after session restarts - **Patrols**: Recurring workflows for queue cleanup and health checks - **Convoys**: Aggregations tracking delivery units like releases ## The 8 Stages of AI Coding Yegge proposes a developer evolution model for AI agent adoption: 1. Zero or near-zero AI usage 2. Coding agent in IDE with permissions enabled 3. Agent in IDE, YOLO mode (permissions off) 4. In IDE, wide agent (code is just for diffs) 5. CLI, single agent, YOLO 6. CLI, multi-agent (3-5 parallel instances) 7. 10+ agents, hand-managed 8. Building your own orchestrator (20-30+ agents) Gas Town targets Stage 6-8 developers. Yegge warns that earlier-stage developers will find the system counterproductive. ## Three Development Loops - **Outer Loop** (days-weeks): Strategic planning, system upgrades, town-level cleanups - **Middle Loop** (hours-days): Agent spawning decisions, Mayor/Polecat coordination, capacity throttling - **Inner Loop** (minutes): Frequent handoffs, task specification, output review ## Current Limitations - Early-stage: freshly ported to Go - Cost: Running 20-30 instances requires substantial compute and API spending - Complexity: Many interdependent roles create a steep learning curve - Reliability: Depends on agent cooperation; some features remain fragile ## References - Original article: https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16dd04 - Gas Town decoded: https://www.alilleybrinker.com/mini/gas-town-decoded/ - Hacker News discussion: https://news.ycombinator.com/item?id=46458936 ## Related - [[Steve Yegge]] - [[AI Agents]] - [[Claude Code]] - [[Beads]] - [[Multi Agent System (MAS)]]