If you’ve ever felt that managing a single AI agent is like herding a very clever cat, try managing a whole team of them. That’s where Gas Town (gt) comes in. It’s the infrastructure for a post-human developer experience, designed to orchestrate multiple AI agents working across distributed workspaces.

What is Gas Town?

Gas Town is a workspace manager for “rigs”—multi-agent environments where agents (called “polecats”) collaborate on a shared codebase. It doesn’t just run an agent; it coordinates their spawning, work distribution, and communication. Think of it as a specialized “OS” for AI-native development.

Core Concepts

  • Rigs: These are the workspaces where the work happens. A rig is usually tied to a specific project or git repository.
  • Beads: Every time an agent starts a session to work on a task, it’s called a “bead.” You can track, resume, or close these beads.
  • Polecats: These are the persistent identities of the agents. While an individual session might be ephemeral, the polecat’s identity and history can persist.
  • The Deacon: A town-level watchdog that keeps everything running smoothly, ensuring agents don’t get stuck and work flows through the system.

How to Use Gas Town

Getting started is simple. Once you’ve initialized your workspace with gt init, the gt command becomes your control center.

Dispatching Work

The most important command is gt sling. This is how you tell an agent to do something.

gt sling "Refactor the authentication module to use JWT"

Communication

Agents in Gas Town can talk to each other and to you via a built-in messaging system.

  • gt mail inbox: Check for messages from agents.
  • gt nudge: Send a quick message to a specific worker.

Monitoring Progress

Because Gas Town is built for distributed work, you need ways to see what’s happening.

  • gt trail: Shows recent agent activity.
  • gt feed: Provides a real-time activity feed.
  • gt status: Gives an overview of the entire “town.”

Why Use It?

The biggest advantage of Gas Town is scale. By decoupling the agent’s identity from the local session and providing a robust communication layer, you can have dozens of agents working in parallel on different parts of a codebase without them stepping on each other’s toes.

It handles the “boring” parts of AI orchestration—state management, session persistence, and inter-process communication—so you can focus on the “what” instead of the “how.”

Silver Umbrella - Technical Learnings