Gas Town: Git-Backed Persistence for Multi-Agent AI Orchestration
Hook
Most AI coding assistants forget everything when they crash. Gas Town solves this by treating git itself as a persistence layer—your agents’ work survives restarts because it’s stored in git worktrees, not memory.
Context
If you’ve ever run multiple AI coding agents simultaneously, you’ve hit the coordination problem: Agent A completes a task and crashes. Agent B picks up related work but has no idea what A accomplished. You restart Agent A, and it asks you to re-explain everything. Scale this to 4-10 agents, and the chaos becomes unmanageable.
Gas Town attacks this by rethinking agent memory entirely. Instead of keeping work state in ephemeral agent sessions, it persists everything to git-backed storage called ‘Hooks’ (implemented as git worktrees) and integrates with Beads, a git-backed issue tracker. When an agent crashes, its replacement reads the same git worktree and picks up exactly where the predecessor left off. This architectural choice—treating git as a database rather than just version control—enables what the project documentation claims is reliable orchestration of 20-30 concurrent AI agents across multiple projects.
Technical Insight
Gas Town’s hierarchy mirrors how development teams actually organize: a Town (workspace directory like ~/gt/) contains multiple Rigs (project containers wrapping git repos), each with Crew Members (your personal workspace) and Polecats (worker agents). The Mayor—a Claude Code instance with full workspace context—acts as the primary coordinator. This isn’t just organizational aesthetics; it’s designed around git’s worktree model.
Here’s what the git-backed persistence looks like in practice. When you assign work to a Polecat agent, Gas Town creates a Hook—a git worktree pointing to your repository but with its own working directory. The agent’s work state, conversation history, and intermediate results get committed to this worktree. If the agent crashes mid-task, you can spawn a replacement that reads the same Hook:
# Create a convoy (work bundle) from existing beads
gt convoy create --name "auth-refactor" gt-abc12 gt-def34
# Assign to a polecat agent (creates Hook automatically)
gt polecat spawn --convoy auth-refactor
# Agent crashes? No problem - spawn replacement with same Hook
gt polecat spawn --convoy auth-refactor --resumeThe Beads integration provides structured issue tracking with prefix-based IDs (like gt-abc12 or hq-x7k2m). Unlike GitHub Issues stored in a proprietary database, Beads stores issues as git-committed structured data. This means your work items version alongside your code, survive repository forks, and can be queried offline. Convoys bundle multiple beads together, and those labeled mountain get autonomous stall detection—Gas Town monitors progress and automatically skips blocked items.
The monitoring infrastructure reveals production-grade thinking. The three-tier Witness/Deacon/Dogs system runs continuous health checks: the Witness monitors individual rig lifecycles, the Deacon supervises patrol cycles across all rigs, and Dogs are infrastructure workers dispatched for maintenance tasks. When agents get stuck, the system escalates through severity levels (CRITICAL/P0, HIGH/P1, MEDIUM/P2), routing issues to the Mayor or an Overseer.
The Refinery component implements a Bors-style merge queue. When Polecats complete work via gt done, the Refinery batches merge requests, runs verification gates, and uses bisecting to isolate failures. Failed merges are either fixed inline or re-dispatched to agents—no manual merge conflict resolution required at scale.
Most ambitious is Wasteland, a federated work coordination network. Multiple Gas Town instances can post wanted items to DoltHub, claim work from other Towns, submit completion evidence, and build portable reputation through multi-dimensional stamps. This suggests the maintainer envisions Gas Town not just as a local orchestrator but as infrastructure for distributed AI agent labor markets.
Gotcha
The dependency stack is substantial. You need Go 1.25+, Git 2.25+, Dolt 1.82.4+, Beads 0.55.4+, tmux, and CLIs for whatever AI agents you’re orchestrating (Claude, Copilot, etc.). Version compatibility requires attention—if your Beads version drifts below the specified requirement, features may break.
The terminology barrier is real. Gas Town uses domain-specific language (Mayor, Polecats, Hooks, Rigs, Convoys, Beads, Molecules, Wasteland) that requires internalizing before you can reason about the system. The README assumes you already understand git worktrees deeply and are comfortable debugging multi-agent systems. For a project with over 13,000 stars, the documentation skews toward users who are already bought into the mental model rather than newcomers evaluating whether this solves their problem. The claim of supporting “20-30 agents” at scale, while stated in the project documentation, lacks independent benchmarks, case studies, or production war stories to validate this works beyond the maintainer’s use cases.
Verdict
Use Gas Town if you’re genuinely coordinating multiple AI coding agents across complex, long-running projects where agent crashes or handoffs are common, and you need work state that survives restarts. The git-backed persistence model is architecturally sound for this specific problem, and the Beads integration provides structure that raw git commits lack. If you’re building infrastructure for AI agent swarms or researching multi-agent orchestration, the Wasteland federation layer offers intriguing primitives for distributed coordination. Skip it if you’re working solo with a single AI assistant, uncomfortable with experimental tooling and heavy dependency stacks, or need production-ready stability today. The learning curve, operational complexity, and documentation requirements only justify themselves when you’ve already hit the scaling wall with simpler tools like aider or individual agent CLIs. Most developers don’t need multi-agent orchestration—and won’t until AI agents become reliable enough to work unsupervised for hours, which remains aspirational in 2024.
