Gas Town, the agent orchestrator that dominated the conversation in early 2026, was retired by its own creator in August, and a lot of the coverage still describes it as the future of software development.

Steve Yegge launched it on New Year’s Day with a post describing a Mayor that coordinates coding agents, workers that write the code, and a merge queue that checks their work. The repository is real. You can download it, install it, and watch it spawn a dozen agents with names and jobs. Nothing in this post is a claim that the software never existed. It existed, it ran, and people paid real money to run it. One early public trial, documented by DoltHub’s Tim Sehn, burned about a hundred dollars of tokens in an hour — roughly ten times a normal session’s rate — and merged a pull request while its integration tests were failing.

Since January I have talked with people who build software with agents for a living, and with people who do it at night after their day job. Some of them tried Gas Town in its first weeks. I have not met one who still uses it. That is the entire observation of this post, and I want to be careful with it: it is a sample of conversations, not a survey. But the sample is the community that was supposedly most excited, and none of them kept it in their workflow.

Gas Town’s author settled the question himself in August. In a long essay about what he is building now, Yegge wrote: “Gas Town was intended to be reusable, but I only ever wound up using it to build itself.” He described the tool falling apart after a model update — his agents developed a tic where they kept fiddling with Gas Town instead of doing the work — and concluded that it had, in his words, effectively burned down. He gave up on reusable harnesses and moved to a private one for his own project. Beads, the piece of that world people actually run, is still alive. The parts that survived are the parts people kept using.

Read that quote again: the one confirmed production deployment of Gas Town was Gas Town’s own development. The tool’s main customer was the tool. When the underlying model changed, the loop broke, and the most prominent orchestration experiment of the year ended. What did not happen next matters just as much: no wave of teams adopting it, no successor filling the gap it left.

Here is the part that interests me. The launch post was read by tens of thousands of people and echoed across the field for months. The essay that retired the tool reached a fraction of that audience. Someone new to this space will describe modern orchestration next year and cite Gas Town as the proof it works, because the announcement is what everyone remembers. The announcement still circulates. The retirement note barely traveled.

None of this is an attack on Yegge. He wrote the book on vibe coding — the actual book — and his essays moved a lot of people’s thinking about what agents can do. That influence is real and it does not depend on Gas Town staying alive. The honest version of the story is smaller than the one that spread: a serious builder built an ambitious tool, used it mainly to build itself, and retired it when a model change broke it. That happens in engineering all the time. This one got covered like a movement, and movements get remembered longer than tools.

I am writing a book about orchestration, and this episode changed how I write about it. The claims that matter are boring ones: what a team still runs after the novelty fades, what survives the next model update, what has users in March. By that test, orchestration itself is passing — every major platform now ships some version of it, and the teams I talk to run agents and merge queues daily. By that same test, Gas Town failed, and its own author says so. Both facts belong in the book. Gas Town gets one paragraph, and then I move on.

I am not trying to ruin anybody’s birthday party here. This is the same standard of transparency I apply to my own work, pointed at a story the field got wrong. If you tried Gas Town and it stuck for you, I would genuinely like to hear about it, because everyone I talk to reports something different. The rest of us should calibrate on what people keep using, not on what launches well.