Yegge Shuts Down Gas Town: Lessons for AI Orchestrators

Steve Yegge—one of the most prominent advocates of autonomous AI agents—shut down Gas Town, his multi-agent orchestrator for writing code.

Several thousand dollars a month on commercial coding agent subscriptions, months of development—and the only thing actually built with this orchestrator was the orchestrator itself. Not a single task beyond it. Researcher Dan

Lu independently described the same problem with similar tools: they abandon tasks unfinished even though the resulting code is usually correct. Gas Town was supposed to build a pipeline of several AI agents that distribute development subtasks among themselves without constant human involvement.

The idea is clear: less manual work means features reach production faster.

In practice, the orchestrator built only itself.

Yegge is not the only one to hit this ceiling: the same conclusion was reached independently by

Dan Lu, while analyzing other “ultra-vibey” orchestrators on his website.

The working approach is a layer around AI agents, with explicit checkpoints after every significant step.

Automatic gates stop the pipeline when tests fail or a contract is violated—the agent cannot continue building on a broken foundation.

A clear protocol separates the roles of agents and tools into distinct contracts instead of relying on one monolithic “orchestrator” that writes code, reviews it, and decides on deployment by itself. MCP is better suited to this than custom wrappers: the boundary between agent and tool becomes explicit and testable, making it verifiable in CI.

A separate issue is the agent’s access to external systems and data.

The broader the range of tasks an orchestrator performs unsupervised, the higher the cost of an access error: committing to the wrong place, deleting the wrong thing, or leaking the wrong secret. LLM & Security Gateway addresses exactly this by controlling which calls to which systems the agent is allowed to make, regardless of what it considers necessary at the moment.