Technical Presentation

GAS TOWN

Kubernetes for AI Agent Orchestration

A Technical Introduction to Multi-Agent Development

Open Source Framework

github.com/steveyegge/gastown

February 2026

Introduction

What is Gas Town?

Multi-Agent Orchestration

Gas Town is an open-source framework that enables developers to coordinate multiple AI agents working together on complex software development tasks.

Created by Steve Yegge

Developed by the legendary engineer and blogger, Gas Town represents a breakthrough in how we think about AI agent collaboration and governance.

Core Purpose

Scale AI Development: Run dozens of agents simultaneously on complex projects

Specialized Roles: Each agent has a specific function in the development pipeline

Cost Efficiency: Optimize agent execution across different model providers

Key Insight

Gas Town treats AI agents like containers in Kubernetes - scheduled, monitored, and scaled as a unified system.

Agent Maturity Model

8 Levels of Agent Development

Level 1

Chatbot

Basic Q&A

Level 2

Assistant

Task Helper

Level 3

Coder

Code Generation

Level 4

Developer

Full Projects

Level 5

Architect

System Design

Level 6

Team Lead

Coordination

Level 7

Manager

Multi-Agent

GAS TOWN

Level 8

Orchestrator

Autonomous System

Gas Town = Level 8 Orchestrator

At Level 8, agents don't just work on tasks - they self-organize, self-heal, and self-optimize . Gas Town represents the pinnacle of agent orchestration, where multiple specialized agents collaborate autonomously under a governance framework.

Autonomous Operation

Self-Healing

Continuous Optimization

Level 8

Orchestrator

The Analogy

The Kubernetes of Agent Orchestration

Kubernetes

Container Orchestration

Manages containers across clusters

Scheduling & Scaling

Auto-scales based on demand

Fault Tolerance

Self-healing, restart failed pods

Gas Town

Agent Orchestration

Coordinates multiple AI agents

Task Scheduling

Assigns work to specialized agents

Agent Health

Monitors & recovers failed agents

Both provide:

Resource Management

Load Balancing

High Availability

Understanding Gas Town

Key Terminology

Agent Roles

Mayor

Central coordinator, manages all agent activities

Polecats

Worker agents that execute development tasks

Witness

Validates and verifies agent outputs

Deacon

Message broker for inter-agent communication

Overseer

Monitors system health and performance

Core Concepts

Beads

Atomic task units that agents work on

Git Worktrees

Isolated environments for parallel development

GUPP

Gas Town Universal Protocol for communication

Refinery

Processing pipeline for agent outputs

Each role has specific responsibilities, creating a governance structure for multi-agent collaboration.

Cost Analysis

Model Pricing Comparison

Price per 1M tokens (USD) - Updated February 2026

Claude Models (Anthropic)

Alternative Models (Cost-Effective)

BEST VALUE

GLM-5

Zhipu AI

Input

$0.80

Output

$2.56

Kimi K2.5

Moonshot

Input

$0.45

Output

$2.20

MiniMax-M2.5

MiniMax

Input

$0.295

Output

$1.20

Savings with GLM-5 vs Sonnet 4.6:

Input: 83% cheaper

Output: 90% cheaper

Strategic Benefits

Why Run Gas Town on Alternative Models?

Massive Cost Savings

40 Agents on Sonnet 4.6

$600/day

40 Agents on GLM-5

$120/day

80% SAVINGS

Reduced Vendor Lock-in

Diversify across multiple providers. If one API has issues, seamlessly switch to alternatives without disrupting your agent workflows.

Claude API Issues

Anthropic API has experienced reliability issues and downtime . Alternative models provide stability and consistent availability for production workloads.

Competitive Performance

GLM-5, Kimi K2.5, and MiniMax deliver near-Claude-level performance at a fraction of the cost, making them viable for most Gas Town use cases.

Configuration Guide

Creating Custom OpenCode Agents

Step 1: List Available Models

All Providers

opencode models

Filter by Provider

opencode models openrouter

opencode models amazon-bedrock

Model ID Format

Format: provider/model-name

openrouter/z-ai/glm-5

openrouter/moonshotai/kimi-k2.5

amazon-bedrock/us.anthropic.claude-sonnet-4-6

Step 2: Create Agent Alias

Command Syntax

gt config agent set "opencode -m "

Examples:

# Claude Sonnet via Bedrock

gt config agent set opencode-sonnet "opencode -m amazon-bedrock/us.anthropic.claude-sonnet-4-6"

# GLM-5 via OpenRouter

gt config agent set opencode-glm5 "opencode -m openrouter/z-ai/glm-5"

# Kimi K2.5 via OpenRouter

gt config agent set opencode-kimi-k25 "opencode -m openrouter/moonshotai/kimi-k2.5"

Note: OpenCode agents need a nudge after starting - they don't auto-prime like Claude Code.

Practical Usage

Using Custom Agents in Gas Town

Start a Crew

Command:

gt crew start myrig stephen --agent opencode-glm5

Launch a crew with a specific agent for all tasks

Set Role Agent

Command:

gt rig settings set myrig role_agents.polecat opencode-kimi-k25

Assign specific agent to a particular role

Set Default

Command:

gt config default-agent opencode-sonnet

Make an agent the default for all operations

Complete Workflow Example

1

Configure Agents

gt config agent set glm5 "opencode -m openrouter/z-ai/glm-5"

gt config agent set kimi "opencode -m openrouter/moonshotai/kimi-k2.5"

2

Assign Roles

gt rig settings set myrig role_agents.mayor glm5

gt rig settings set myrig role_agents.polecat kimi

3

Launch Crew

gt crew start myrig stephen

4

Monitor & Scale

gt crew status myrig

gt crew scale myrig --agents 20

Cost-Optimized Multi-Agent System

Get Started

Resources & Community

GitHub Demo

Complete working example of Gas Town with sample rigs, configurations, and agent definitions.

Repository:

github.com/stephenVertex/gastown-demo

Includes sample configurations

Gas Town Hall

Official community hub for documentation, tutorials, best practices, and user-contributed rigs.

Website:

gastownhall.ai

Documentation & guides

Discord Community

Join the active community for real-time support, troubleshooting, feature discussions, and collaboration.

Invite Link:

discord.gg/xHpUGUzZp2

Active community support

Ready to orchestrate your agents?

Star the Repo

Join Discord

Share Your Rigs

THANK YOU

Gas Town: Kubernetes for AI Agent Orchestration

stephen.barr@devfactory.com

Questions? Reach out or join the Gas Town Discord community.

github.com/steveyegge/gastown

gastownhall.ai