I Let 30 AI Agents Loose in My Repo (Gas Town)

The speaker introduces Gastown as a paradigm shift in software development where Claude Code is transformed from a single assistant into a team of 30 to 50 agents. Instead of a simple chat interface, the system functions as a factory that decomposes features and assigns parallel workers to ship code at scale. Key technical benefits mentioned include automatic code merging, Git persistence, and the ability to continue running after a crash. This section highlights how the tool can potentially 10x developer output while cautioning that it requires proper management. It sets the stage for a technical deep dive into the orchestrator's architecture and setup process.

This segment explains that Gastown is an open-source orchestrator sitting on top of models like Claude Code rather than being an AI model itself. The speaker details the hierarchy starting with the 'Mayor' who breaks goals into 'Beads,' which are then grouped into 'Convoys' for the 'Polecat' workers to execute. A major advantage discussed is the use of Git branches to isolate work and queue merges, preventing the context drift common in 'vibe coding.' The speaker describes this as 'horizontal scaling for cognition,' contrasting it with the messiness of manual, linear AI orchestration. The section concludes with the speaker transitioning into a real-world test to see if this complex system actually delivers on its promises.

The speaker selects a simple Go-based to-do application to test Gastown's ability to implement a multi-layered feature like JWT authentication. Adding authentication is presented as a high-friction task that typically involves changes to databases, routes, middleware, and tests. A single prompt is given to the 'Mayor' to add login/register endpoints, use SQLite, protect routes, and create a Docker file. The speaker mentions that while the initial environment setup was a 'pain' involving Brew installs and troubleshooting, the system eventually ran smoothly. This part illustrates the complexity of the task being offloaded to the AI and the preparation required for such a sophisticated workflow.

This section shows the 'Mayor' decomposing the prompt into separate 'Beads' for schema creation, middleware, and frontend forms. Multiple 'Polecat' workers are seen swarming the repository in parallel, accomplishing in minutes what would take hours of sequential manual prompting with Claude Code. The speaker reviews the GitHub repository to find that the AI successfully added all required files, passed integration tests, and issued valid tokens. This is framed as a shift from 'managing every AI step' to 'cognitive offloading' where the system holds the entire project plan in its state. The speaker confirms that the resulting application is fully functional, with the to-do list intact and all new features working correctly.

The final segment provides a balanced look at the pros and cons of using a multi-agent system like Gastown. Pros include immense scalability and Git-based persistence, which prevents data loss during crashes; however, significant cons are 'oversight fatigue' and high API costs. The speaker warns that agents can still conflict or over-engineer code, potentially 'wrecking' a repository if left completely unattended. Gastown is presented as an early-stage tool that points toward a future where AI isn't just an assistant but a fundamental change to the entire development lifecycle. The video concludes with an invitation for viewers to decide if they want to experiment with these multi-agent systems now or wait until they become the industry standard.