Using one AI coding agent like Claude Code feels normal by now. But the moment you spin up several, you hit a wall fast. A dropped session wipes out whatever the agent was doing, and you can't even see who is working on what.
A handful of open-source projects are tackling this head-on. Here's a look at how Gas Town structures its answer, and how it differs from three projects in the same space: Paperclip, Vibe Kanban, and Multica.
What Gas Town Is Going For#
Gas Town is a multi-agent orchestration system for AI coding agents. It is designed from the start to run many agents at once, not just one.
The core philosophy comes down to a single idea: an agent should never lose its work state, even after a restart. To get there, context lives in a Git-based persistent store instead of memory, handled through two concepts called Hooks and Beads.
The Four Characters#
Gas Town's structure reads like a small town, right down to the names. There are four roles.
- Mayor — The main AI coordinator. You talk to the Mayor to hand out work.
- Polecats — The worker agents that actually do the tasks.
- Hooks — Persistent work storage built on Git worktrees. This is where an agent's work is kept.
- Convoy & Beads — Each task is tracked as stateful data called a Bead, and multiple tasks are bundled and tracked as a Convoy.
You don't address the workers one by one. You only talk to the Mayor, and the Mayor drives everyone else.
Three Features That Stand Out#
1) Git Worktree-Based State Persistence#
Even if an agent's session drops or restarts, its in-progress work is stored in the Git-based Hooks structure. Bring it back up and it picks up exactly where it left off.
This is the most painful spot when you run multiple agents. If every dropped session means explaining everything from scratch, there is no point in running several at all.
2) The Mayor, a Big-Picture Coordinator#
You don't have to decide which agent gets which job. Give the Mayor the overall goal, and the Mayor distributes work to the agents below it. Gas Town calls this dispatch action a Sling.
The person who sets the goal is separated from the coordinator who breaks it down and hands it out. You get to step back from the PM role.
3) TUI Visibility and Nudges#
Run gt feed and you get a terminal UI showing the current state of dozens of agents at a glance. States come in flavors like Working, Stalled, and Zombie.
When you spot a stuck agent, you can nudge it back into motion. It gives you a place to catch an agent that has looped or died in place and poke it yourself.
How It Differs from Paperclip, Vibe Kanban, and Multica#
There are other open-source projects working on the same problem. All four deal with "how do you run multiple agents," but each puts its weight in a different place.
| Tool | Focus | Approach |
|---|---|---|
| Paperclip | Organizational control | Treats agents as a "company" and focuses on org charts, roles, budget control, and policy management |
| Vibe Kanban | Kanban and review | Removes the human planning and review bottleneck, isolating work in Git worktrees to provide previews |
| Gas Town | Session persistence and autonomous dispatch | A hierarchical agent army where a main AI drives sub-agents through the Mayor model |
| Multica | Issue tracker and work queue | Humans and agents share one issue board, and a local daemon picks up queued tasks and runs them |
Paperclip cares about controlling who is allowed to do what. Vibe Kanban focuses on making the point where a human reviews as comfortable as possible.
Gas Town, unlike those two, builds a hierarchy among the agents themselves. Because a main AI coordinator drives the sub-agents, the point where a human steps in sits at the very top.
Multica Starts from the Issue Tracker#
Multica is an open-source, self-hostable project whose pitch is "make humans and AI agents work as one team." Unlike the other three, its starting point isn't the agent. It's the issue tracker.
Just as you'd assign a person to an issue, you assign an agent as the assignee. A daemon running on your local machine then picks the task off the queue and launches a CLI like Claude Code, Codex, Gemini, or Copilot. Results land as issue comments and status changes, and when a linked GitHub PR merges, the issue automatically moves to done.
It also has squads, which group agents and route work to a leader, and autopilots, which create tasks automatically from a schedule or a webhook. Where Gas Town's Mayor divides work inside a conversation, Multica makes the issue board itself the work queue, so a human can jump in from the same screen at any time.
Wrapping Up#
The four tools each pick a different axis. Paperclip brings organizational control, Vibe Kanban brings visual review, Gas Town brings automatic dispatch and health checks through the Mayor, and Multica brings a work queue built on an issue tracker.
When choosing a multi-agent tool, don't start with "which one is best." Start with which axis hurts you the most. If dropped sessions are what's killing you, that maps exactly onto the problem Gas Town solves. If you want agent work recorded issue by issue, Multica is the closer fit.
Don't settle for a relationship that won't let you be yourself.
— Oprah Winfrey
