What Is Gastown? How Steve Yegge's AI Coding Agents Are Changing Software Development

AI-generated, human-reviewed.

AI orchestrators like Gastown are setting the stage for a new era in software development, where teams of intelligent agents collaborate to build complex software with minimal human supervision. On Intelligent Machines, guest Steve Yegge shared how his open-source project, Gastown, brings a radically different approach to coding with AI—and what that means for developers and enterprises today.

What Is Gastown? The AI Agent Coding "Factory"

Gastown is not just another coding assistant. According to Steve Yegge on this week's Intelligent Machines episode, it's an orchestrator—a system where one AI agent manages a team of other AI agents, each specializing in different roles. Rather than simply generating code on demand, Gastown allows a "mayor" agent to oversee a "town" populated by coders, reviewers, and supervisors, all powered by Claude (Anthropic's AI) and interconnected through a coordinated workflow.

Gastown automates and delegates tasks to generate, test, and review code, much like how a real software team would operate.

For those new to the concept, an "agent" in this context is an AI instance given a clear set of rules and responsibilities, and can operate semi-independently. Gastown assigns agents to specialized tasks—one might focus on designing architecture, another on code review, while others document or test the software, working together towards a shared goal.

The Key Innovation: Orchestration and Memory

Yegge explained that the breakthrough behind Gastown and similar systems is orchestration—having agents that coordinate the efforts of other agents.

A recurring challenge with current AI models is memory: because they don't retain context over time, they can easily lose track of complex, multi-stage software projects. Gastown introduces a solution called "beads," a lightweight, structured issue tracker that acts as an external memory for the agents. Each agent can reference, update, and share these "beads," ensuring the team doesn't lose continuity—even when individual prompts end or agents swap roles.

This approach not only maximizes productivity, but it allows for a persistent knowledge graph of tasks, ideas, and ongoing work.

Why Gastown Is Not for Everyone (Yet)

While AI agent factories sound enticing, Yegge and the hosts repeatedly cautioned that Gastown is not consumer-ready. As of this episode, it remains experimental—best suited for advanced users, researchers, and enterprises willing to navigate complexity in exchange for cutting-edge capabilities.

Yegge noted that Gastown requires substantial manual oversight, frequent corrections, and a willingness to operate "as if you're managing a very fast, very junior dev team." Technical users also need to manage risk, as improperly controlled agents could unintentionally expose credentials or make other costly mistakes.

Anyone interested in these tools should approach with caution, and be aware that the AI-generated output still requires stringent review and quality assurance.

How Gastown Is Used Today

Despite these challenges, Gastown is already being adopted for real work—including by some Fortune 100 companies looking for new automation opportunities.

Gastown supports the ability to:

• Parallelize research and coding tasks
• Review dozens of pull requests (PRs) per day using multiple AI-powered "crew" agents
• Persist knowledge and instructions using the beads system (currently implemented with JSON and SQLite, with plans to migrate to more scalable databases)
• "Federate" multiple towns, enabling a higher level of distributed work

Users are finding that, with proper care, this setup can lead to surprisingly productive results and faster iteration cycles.

Key Takeaways

• Gastown exemplifies the future of AI agent orchestration in coding, letting teams of agents work together with minimal human direction.
• Beads provide the crucial memory and task management that allow agent systems to track, persist, and coordinate work over time.
• Adoption is currently best for advanced users: Gastown remains risky and requires expert supervision.
• The productivity gains are real: businesses are already finding practical use cases, especially for automating repetitive or research-heavy tasks.
• Yegge and the hosts advise caution: keep security, manual review, and checks at the forefront, and don't hand sensitive accounts or information to agents without proper safeguards.

The Bottom Line

Gastown and similar AI agent orchestrators represent a dramatic shift in how software may be built in the coming years. While the tools are still raw and demand careful oversight, they showcase how coordinated AI teams could soon automate significant portions of the development process—reshaping the work of engineers, project managers, and entire organizations.

To follow the latest on this topic, including Steve Yegge's future writing and developments in multi-agent AI systems, stay tuned to Intelligent Machines.

Listen or subscribe to Intelligent Machines episode 856:
https://twit.tv/shows/intelligent-machines/episodes/856