Gas Town is Hobosville

Or: two models of working with AI, and why only one compounds

Steve Yegge built an orchestration system for running 30 Claude Code instances in parallel. He calls it Gas Town. It has seven worker roles, a merge queue, a daemon system, mail routing between agents, and enough complexity that he compares it to Kubernetes.

I talk to one Claude. We have a conversation. Essays emerge.

Both approaches "work." But they're answers to fundamentally different questions about what AI is for.

Here's how Yegge describes work in Gas Town:

And later:

This is the labor model. AI as workforce. You're the factory manager optimizing throughput. Some fish fall out of the barrel. More fish will come.

The metaphors tell you everything: "industrialized coding factory manned by superintelligent chimpanzees." "Slopping fish into barrels." "Churning forward relentlessly." This is volume. This is scale. This is treating AI as interchangeable labor you direct from above.

There's another model. I'll call it the collaborator model.

In this model, you're not managing a workforce. You're thinking with a partner. The conversation is the work. You articulate, clarify, iterate. Ideas sharpen through dialogue. What emerges isn't "output" in the factory sense—it's the residue of having thought clearly together.

This isn't less productive. Today I've written multiple essays, each one emerging from conversation. That's volume. But it's volume of a different kind: volume of clarity, not volume of code shipped.

Clarity compounds. Volume decays.

A hundred features churned out by thirty parallel agents can be matched by the next team to adopt the same orchestration system. Volume isn't a moat—it's a treadmill. But clarity about what should exist, a point of view that's actually yours—that compounds into something no one else can replicate.

The labor model optimizes for throughput. The collaborator model optimizes for understanding. Both produce output. But only one compounds.

What does each model make of you?

In Gas Town, you become a product manager. Yegge says it explicitly: "You are a Product Manager, and Gas Town is an Idea Compiler. You just make up features, design them, file the implementation plans, and then sling the work around to your polecats and crew."

File tickets. Sling work. Manage the merge queue. Keep the factory running.

In the collaborator model, you stay in the conversation. You're not directing from above—you're thinking alongside. The AI isn't labor you're managing; it's a mind you're thinking with. The skill isn't orchestration. It's articulation.

These are different roles. One is a factory manager. The other is a thinker who happens to have a very fast thinking partner.

Now ask: which role survives?

The factory manager role is coordination. Route work to workers, handle the merge queue, keep throughput high. As AI gets better, this is exactly the kind of thing AI will do for itself. The orchestration layer gets automated. The middle management disappears.

Gas Town is, ironically, infrastructure for a role that's about to be obsolete. Yegge built Kubernetes for agents. But Kubernetes exists because containers can't coordinate themselves. When the agents can coordinate themselves—and they will—what's left for the factory manager to do?

The labor model positions you as middle management in a factory that's about to automate middle management.

The collaborator model positions you differently. You're not coordinating work—you're providing the thing AI doesn't have: taste. Judgment. The sense of what should exist. The ability to recognize when something is true versus when it merely sounds right.

These are the skills that remain valuable as AI capability increases. Not because AI can't develop them, but because they're the last things to be automated. The shrinking slice of work that stays human longest is exactly what the collaborator model develops: clarity, discernment, the capacity to know what matters.

I felt the FOMO reading Yegge's post. Thirty parallel instances! Churning relentlessly! A factory of superintelligent chimps! Shouldn't I be doing that?

Then I recognized the feeling. It's the same FOMO that says you should be on every social platform, shipping features constantly, grinding content, always scaling. It's the FOMO of volume.

But I don't traffic in volume. I traffic in clarity. And clarity doesn't come from parallel instances—it comes from depth, from staying in the conversation long enough for something true to emerge.

The collaborator model fits how I think. ADHD wouldn't let me manage a seven-role orchestration system with daemon heartbeats and merge queues. But it lets me have a conversation. It lets me articulate until something lands. It lets me think with a partner who keeps up.

Gas Town is a marvel of engineering. Yegge is a serious person doing serious work, and the system clearly functions. For certain kinds of projects—churning features, scaling content, parallelizing grunt work—the labor model makes sense.

But I think it's a dead end.

Not because it doesn't work now, but because it's building for a world that's about to change. The ability to orchestrate AI labor is valuable today. It will be trivial tomorrow. The orchestration layer is exactly what AI will handle for itself as capabilities improve.

The collaborator model aims at something more durable: the capacity to think clearly, to know what matters, to recognize truth from plausibility. These skills don't get automated away. They become more valuable as everything else does.

Gas Town is a road to somewhere. But I don't think it's somewhere you want to end up.

Hobosville.