How to Command 30 AI Agents: Scaling Coding Productivity 10x with Gastown

The era of asking a simple chatbot for code and waiting for an answer is over. While Claude Code or GitHub Copilot are excellent assistants, they are too slow and linear for modifying enterprise systems where hundreds of files are intertwined. Relying on a single AI that loses context and wanders during long sessions is now nothing more than a productivity bottleneck.

Now is the era of Agent Orchestration. Gastown, an open-source project proposed by Steve Yegge, aims for a "coding factory" system that runs more than 30 AI agents simultaneously to decompose features and build them in parallel. You must now stop being a coder and become an orchestrator commanding an AI legion.

Gastown Architecture: Perfect Results from Unreliable Workers

Gastown is not just a simple AI wrapper. It adopts distributed computing principles to solve the instability of individual agents through system structure. The key is to clearly separate roles and break down tasks into atomic units.

• Mayor: The central coordinator of the entire system. It analyzes the developer's goals and decomposes them into small units of work called Beads.
• Bead: Persistent issue data backed by Git. This is the core mechanism that preserves work history even if an agent session crashes.
• Polecat: One-time workers that actually modify code in independent Git worktrees. They vanish immediately after the task is completed to prevent code conflicts.

The core of this structure is escaping from Vibe Coding (coding by gut feeling). Gastown physically prevents agents from committing directly to the main branch through a PreToolUse hook system. All code is written in separate feature branches and must pass the Refinery's rigorous test validation before being integrated.

Agent Deployment Strategy by Project Scale

Deploying dozens of agents for every task is a waste of API costs. You must distribute firepower appropriately according to the complexity of the task.

When running 30 agents during a large-scale migration, a task that would take 6 hours of manual work is shortened to around 20 minutes. However, this assumes optimization of model allocation. Maximize cost efficiency by assigning high-performance models like Claude 3.5 Sonnet to the Mayor in charge of design, and cost-effective models like Gemini to the Polecats performing simple implementation or testing.

Practical Application: Guide to Building a JWT Auth System

Suppose you are adding JWT authentication to a Go-based application. Here is how you move the legion with a single line of command:

1. Access the Mayor: Start the session with the gt mayor attach command.
2. Set the Goal: "Build a JWT system for the Go project. Include DB schema changes, login API, middleware, and integration tests."
3. Orchestration: "Do not write it yourself; decompose the tasks and delegate them to Polecats. Perform reviews with the Refinery."
4. Monitoring: Check the progress of each agent in real-time with gt convoy list.
5. Review: Approve the final result merged by the Refinery after checking it with gt status.

If things aren't running smoothly, check your environment first. It is essential to ensure that your Dolt version is 1.82.4 or higher. Older versions cause Git database synchronization errors, leading to task conflicts between agents. Also, if there are issues running the daemon, verify that tmux -V is version 3.0 or higher, then run gt doctor --fix to initialize the environment.

The New Role of the Senior Engineer: System Architect

Operating 30 AI agents simultaneously means you are no longer typing. Now, an engineer's true skill lies in how sophisticatedly they record architectural decisions in instruction documents like CLAUDE.md.

Agents are great assistants, but without proper management, they are like super-intelligent chimpanzees that can paralyze a system. Always run them in a separate experimental Rig and set API cost limits. To reduce the fatigue of reviewing dozens of PRs individually, deploy additional "Sheriff" agents to act as PR gatekeepers, filtering out syntax errors and failed tests in the first pass. Your software factory is now ready for operation.