Steve Yegge's Gas Town Series — Agent Orchestration and the Dark Factory Future
Source: Steve Yegge's Medium blog series, January–April 2026
Canonical URL: https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16dd04
Reading Order: 01: Welcome to Gas Town → 02: Emergency User Manual → 03: Welcome to the Wasteland → 04: From Clown Show to v1.0 → 05: Welcome to Gas City
Post-by-Post Summary
01 — Welcome to Gas Town (Jan 1, 2026)
Gas Town is introduced as "a new take on the IDE for 2026" — an orchestrator that manages 20–30 concurrent Claude Code instances productively. Yegge had been predicting this since March 2025 ("Revenge of the Junior Developer") and built four failed precursors before landing on the Go-based Gas Town.
Core thesis: Claude Code is just a building block. The real product is the workflow layer on top. Gas Town = Kubernetes for agents.
Key architectural roles introduced:
- **Mayor**: The user's primary interface — a concierge/chief-of-staff that handles reading all agent output so the human doesn't have to
- **Refinery**: Decomposes vague epics into well-specified bead sequences for polecats
- **Witness**: Quality auditor watching all workers
- **Deacon**: Patrol daemon keeping workers from going stale
- **Polecats**: Ephemeral, unmonitored workers; given a bead and let loose
- **Crew**: Named, long-lived per-Rig agents for design/back-and-forth work
- **Boot**: Daemon handling heartbeats so the Deacon isn't interrupted
Core principles:
- **GUPP** (Gastown Universal Propulsion Principle): If there is work on your hook, you MUST run it
- **Beads as Universal Data Plane**: All work, coordination, and messaging flows through Beads
- **Zero Framework Cognition**: Agents work with Gas Town without special training
The 8 Stages of Dev Evolution To AI are introduced as a framework for where practitioners are on the adoption curve (Stage 8 = building your own orchestrator).
03 — Welcome to the Wasteland: A Thousand Gas Towns (Mar 4, 2026)
The Wasteland is a federated trust network linking thousands of Gas Towns to work on shared public projects.
Core insight: 100x token spend increases require federating 100x users. The Wasteland's Wanted Board lets anyone post work; rigs claim it, complete it, and receive stamps — multi-dimensional attestations (quality, reliability, creativity) from validators with trust levels.
The yearbook rule: You can't stamp your own work. Reputation is what others write about you, not what you claim about yourself.
Trust ladder: Registered participant → Contributor → Maintainer. Federated wastelands (team, company, university, open source) share the same schema; your reputation portable across all of them.
Underlying protocol: Git's fork/merge-pull model, backed by Dolt (Git for SQL databases). All models already know Git, so adoption is frictionless.
The Wasteland is explicitly framed as an RPG-like reputation economy for AI-augmented work. Pre-seeded with GitHub's top ~10k contributors' data.
05 — Welcome to Gas City (Apr 24, 2026)
Gas City tears apart Gas Town and rebuilds it as composable declarative building blocks called "packs". The Gas Town pack ships as the default, making Gas City a drop-in replacement for existing users.
Light Factory framing: Dark factories (autonomous agent teams) should maximize observability. Gas City is "the Light Factory" — all workers visible and addressable, with the polecats in the back rooms being the only ones you normally don't look in on.
The 11 Stages of AI Adoption extends the original 8:
- Level 8: Building your own orchestrator
- Level 9: Deploying your first agent pack (a "garden you tend")
- Level 10: Multiple packs, each managing a domain, human as control plane
- Level 11 — Factory Builder: Crew managing subsets of packs → full custom orchestrator
Multi-agent reliability principle: Never deploy a single agent for real business processes. Two agents watching each other (like two hash functions) dramatically decrease failure probability. This is why dark factories beat single-agent setups.
Escape From SaaS Mountain: Non-technical staff rebuilt a $30k/year SaaS in Gas Town. The implication: you only need to build the 20% of a SaaS you actually use. SaaS Mountain is cracking from the bottom up.
Gas City is MIT-licensed, community-driven, with a Discord of thousands.
Relation to Kelly/Dark Factory Methodology
| Yegge Concept | Kelly Equivalent |
|---|---|
| Mayor (reading/filtering for human) | Router/Manager agent (context filtering) |
| Polecats (ephemeral workers) | Sub-agents spawned for tasks |
| Crew (long-lived named agents) | Persistent agent sessions |
| GUPP (work persistence via hooks) | Kelly's autonomous continuation |
| Beads (git-backed work items) | Kelly's work-item tracking + TEA audit |
| Witness (quality gate) | Kelly's Gate pattern |
| Wasteland stamps (reputation) | Kelly's autonomous company reputation |
| MEOW (work as currency) | Kelly's factory throughput model |
| Multi-agent adversarial review | Kelly's Angry Mob / 5-agent verdict |
| Light Factory (observability maximized) | Kelly's transparent factory运作 |
Yegge independently arrived at most of Kelly's key patterns through pure intuition and hands-on experimentation with live multi-agent systems. The convergence is significant.