Gas Town GA — $100/Hour to Turn One Dev Into a Team

Steve Yegge’s Gas Town — the multi-agent orchestrator that coordinates 20–30 parallel Claude Code agents through a Mayor, worker Polecats, and a Refinery merge queue — went generally available on Kilo Cloud on May 19, 2026. The pitch is “turn one developer into an engineering team.” The price tag is approximately $100 per hour in model tokens at full fleet capacity. That number is not a footnote. It is the entire decision.

TL;DR

• What: Gas Town by Kilo hit GA — cloud-hosted multi-agent orchestrator with 20–30 parallel Claude Code agents, one-click deploy
• Cost: ~$100/hour in Claude tokens at full capacity, plus Kilo infrastructure surcharge (exact markup not public)
• Novel piece: The Wasteland — a Dolt-backed federated task coordination protocol with reputation stamps
• Verdict: Proof that multi-agent orchestration is a real category. Not a tool most teams should deploy next sprint.

Gas Town GA — What Happened

Gas Town has existed as an open-source project since Yegge announced it in January 2026. The GA milestone is specifically about the cloud-hosted version on Kilo Cloud: managed infrastructure, elastic Polecat scaling, auto-recovery, and access to 500+ models through the Kilo Gateway — all behind a single API with consolidated billing. No more wrangling tmux sessions and per-agent API keys.

The architecture is genuinely distinct from anything else in the orchestration space. A Mayor agent decomposes work into parallelizable tasks. Polecat worker agents each operate in isolated git branches, executing independently. The Refinery manages the merge queue and handles conflict resolution when those branches converge. Every piece of work state persists in git-backed hooks, which means if an agent crashes mid-task, the work survives the restart. That last claim is architecturally sound given the git-branch isolation — each Polecat’s state is the branch itself — though no one has published formal results from multi-day convoy runs yet.

Launching alongside GA is hosted support for the Wasteland, Yegge’s federated task coordination protocol. Built on Dolt — a SQL database with Git semantics — the Wasteland lets developers and agents post work to a shared Wanted Board, claim tasks, submit evidence of completion, and get validated through a multi-dimensional stamp system covering quality, reliability, and creativity. Each stamp traces back to actual work artifacts, so reputation accumulates from evidence rather than self-reporting. The whole thing runs through DoltHub pull requests, making the audit trail inspectable. It is an ambitious governance primitive for agent-driven development, though it is still early enough that calling it production-proven would be generous.

The open-source Gas Town repo at gastownhall/gastown remains available. Kilo Cloud is a managed hosting option, not a replacement. You can still self-host — you just lose the elastic scaling, auto-recovery, and consolidated billing.

Why This Matters

The $100/hour number needs context, because it is both more and less alarming than it sounds. That figure represents full fleet capacity — 20–30 parallel Claude Code agents hammering tokens simultaneously. You are not obligated to run a full fleet. But the system is designed for exactly that scale, and Yegge has been explicit that Gas Town targets developers at Stage 7 or Stage 8 of his Developer Evolution model — people who are already comfortable running 10+ parallel agent sessions daily. If you are not in that category, Gas Town is not built for you, and Yegge says so himself.

The economics break down into three layers. First, raw token costs: Claude is expensive at volume, and 20–30 concurrent agents burn through tokens at industrial rates. Second, Kilo Gateway consolidation: routing 500+ models through a single billing interface removes API key management overhead but does not meaningfully reduce the per-token price. Third, Kilo Cloud infrastructure: the exact surcharge Kilo adds on top of raw token costs is not publicly documented. You are paying for managed orchestration on top of already expensive model inference, and the total bill is opaque until you are running it.

Compare this to framework-based orchestrators like CrewAI, OpenAI Agents SDK, or AWS Strands. Those systems give you the abstraction layer for defining agent roles and workflows, but they do not prescribe the kind of rigid role hierarchy Gas Town enforces. CrewAI lets you compose crews flexibly. Strands gives you tool-use primitives on Bedrock. Gas Town says: here is a Mayor, here are Polecats, here is a Refinery, and the merge queue is not optional. That rigidity is a feature if your problem matches the pattern — parallel code generation tasks that must converge into a single codebase — and a constraint if it does not.

The deeper differentiator is persistence. Most agent frameworks treat sessions as ephemeral. Gas Town’s git-backed architecture means each Polecat’s work is a real branch with real commits. If Kilo Cloud goes down mid-run, the branches still exist. If you need to audit what Agent 17 did at 3 AM, you read the git log. This is architecturally closer to a distributed build system than to a chatbot orchestrator, and it is the single strongest argument for Gas Town over the competition.

Kilo Cloud is the only cloud-hosted version of Gas Town. If Kilo’s pricing or reliability changes, migrating back to self-hosted is non-trivial — you lose elastic scaling, auto-recovery, and the consolidated billing layer. For teams considering production adoption, this is a real lock-in surface worth evaluating before committing.

The Wasteland is the piece that could matter most long-term — or not at all. A federated task marketplace where agents and developers coordinate through validated reputation is a genuinely new idea in developer tooling. But it requires network effects to function. A Wanted Board with three participants is a to-do list. A Wanted Board with three thousand is an economy. Whether the Wasteland crosses that threshold depends entirely on adoption velocity, and right now it is launching cold.

The multi-agent long-running session stack we profiled earlier this year outlined the architectural requirements for systems like this: persistent state, branch isolation, merge governance, crash recovery. Gas Town checks every box. The question was never whether the architecture was sound. The question is whether the token economics make it viable outside a narrow band of high-output, high-budget engineering operations.

The Take

Gas Town GA is the clearest signal yet that multi-agent orchestration is a real product category, not a research curiosity. The Mayor/Polecat/Refinery architecture is genuinely novel — it is not another wrapper around a chat API, and it is not a framework you have to wire together yourself. The git-backed persistence model solves a problem that most orchestrators ignore entirely. The Wasteland is an ambitious bet on federated agent governance that could become important if it reaches critical mass.

But I would not deploy it next sprint. The $100/hour token burn at full capacity means Gas Town only makes economic sense for teams where a single developer’s parallel output can replace multiple hires — and where the tasks decompose cleanly into independent, mergeable branches. That is a narrower use case than the marketing suggests. If your work involves heavy cross-file dependencies, shared state mutations, or design decisions that require human judgment at each step, 30 parallel Polecats will generate 30 parallel merge conflicts.

The AWS Strands lock-in analysis applies here in a different form: Kilo is not locking you into a cloud provider’s model ecosystem, but it is the sole cloud host for a system that is painful to self-manage. That is a dependency worth naming before you build workflows around it.

My recommendation: treat Gas Town GA as a category validator. Watch the Wasteland adoption curve. Run the open-source version on a contained project to test whether your workloads actually parallelize the way the architecture assumes. Do not sign up for $100/hour cloud billing until you have evidence that the output justifies the spend. The architecture is real. The economics are unforgiving.