Gas Town: What Kubernetes for AI Coding Agents Actually Looks Like
Single AI coding agents are already reshaping how software gets built. But a harder question is emerging: What happens when one agent isn’t enough?
Steve Yegge thinks he has the answer. The veteran engineer — 40+ years at Amazon, Google and Sourcegraph — spent the second half of 2025 building Gas Town, an open-source orchestration system that coordinates 20 to 30 Claude Code instances working in parallel on the same codebase. He describes it as “Kubernetes for AI coding agents.” The comparison isn’t just marketing. It’s architecturally accurate.
From Single Agent to Factory Floor
Most developers using AI coding tools today work with a single agent. Yegge maps the progression across eight stages, from basic code completion to manually running 10+ parallel agents. Gas Town targets top developers already juggling multiple agents who need a single point of contact for their management.
The core problem is familiar to anyone who’s run distributed systems: Coordination. When multiple agents generate code against the same repo simultaneously, you get merge conflicts, duplicated work, and agents that stop when their context window fills up.
The Architecture DevOps Teams Will Recognize
Gas Town’s structure maps surprisingly well to Kubernetes. Both coordinate unreliable workers toward a goal. Both have a control plane that monitors execution nodes, with local agents monitoring ephemeral workers. Both reconcile against a persistent source of truth.
The key difference: Kubernetes asks “Is it running?” Gas Town asks, “Is it done?”
Seven specialized roles divide the work. The Mayor is your primary interface. Polecats are ephemeral workers that execute tasks and produce merge requests. The Refinery manages a merge queue, so parallel work doesn’t collide. The Witness monitors worker health. The Deacon runs patrol loops. Dogs handle maintenance. And your Crew are persistent agents for collaborative design work.
All state lives in Beads — Yegge’s Git-backed issue tracking system that serves as both data plane and control plane. Agent identities, work assignments, and orchestration state all persist in Git. Sessions are ephemeral cattle. Agents are persistent identities.
Durable Workflows Without Deterministic Replay
Gas Town’s most interesting technical concept is what Yegge calls Nondeterministic Idempotence. Unlike Temporal’s deterministic, durable replay, Gas Town achieves workflow durability through a different mechanism.
All work is expressed as “molecules” — chained sequences of small tasks stored as Beads in Git. Each step has clear acceptance criteria. If an agent crashes mid-step, the next session picks up where it left off. The path is nondeterministic — agents might take different approaches each time — but the outcome converges because the workflow definition is persistent and the acceptance criteria are explicit.
This makes Gas Town workflows survive agent crashes, context window exhaustion, and session restarts. You keep throwing agents at the work, and it eventually finishes.
According to Mitch Ashley, VP and practice lead, software lifecycle engineering, The Futurum Group, “Gas Town is a great example of moving to orchestrated agent execution. When development moves from single agents to coordinated swarms, someone owns the control plane. Yegge’s Kubernetes analogy is architecturally sound: both coordinate unreliable workers toward persistent goals, separating control logic from ephemeral execution.”
Ashley continues, “DevOps teams face execution obligations. CI/CD pipelines, merge strategies, and testing must accommodate agent swarms, not individual developers. The merge queue problem becomes universal as multi-agent development reaches production. The orchestration layer determines workflow ownership.”
The Uncomfortable Reality
Gas Town works. But Yegge is remarkably honest about what “works” means at this stage.
It’s expensive. Running 12 to 30 parallel agents burns through API credits fast — one early adopter reported $100 per hour. Work can be chaotic. Bugs get fixed multiple times by different agents. Designs occasionally go missing. And it demands expertise — Yegge warns that developers who don’t already run multiple parallel CLI agents daily will find it counterproductive.
But the throughput is real. The system can run overnight unattended. Workflow durability means nothing is permanently lost.
Why DevOps Teams Should Pay Attention
Gartner reported a 1,445% surge in multi-agent system inquiries between Q1 2024 and Q2 2025. Anthropic lists “mastering multi-agent coordination” as the top organizational priority for 2026. The industry is converging on the same conclusion Yegge reached months earlier: the next step after AI writes your code is orchestrating the AI that writes your code.
For DevOps teams, the implications are direct. CI/CD pipelines, merge strategies, and testing frameworks must accommodate code from agent swarms rather than individual developers. The merge queue problem — which Gas Town addresses with its Refinery — will become a universal concern.
Gas Town is rough. It’s 20 days old and admittedly “100% vibe coded.” But it’s the clearest picture we have of what orchestrated AI development actually looks like—messy, expensive, fast, and focused on the future.
