From Prompt Experiments to Operable Agent Systems: Gas Town on Microsoft's AI Stack

Exploring what a multi-agent development system like Gas Town would look like using Microsoft Foundry and GitHub Copilot's pro-code tooling.

AI Agent Development

Introduction

In January 2026, Steve Yegge introduced Gas Town, an experimental multi-agent development system built around a simple but radical idea: instead of using AI as a coding assistant, use AI as a coordinated workforce.

Gas Town runs 20–30 agents in parallel. Each agent has a role. The developer acts as an Overseer, not a typist — coordinating work rather than writing individual functions.

The original implementation is built on Claude Code and a substantial custom orchestration layer written in Go. But the architectural pattern itself — hierarchical multi-agent coordination with persistent state — is not tied to any specific runtime.

This article explores what a Gas Town-style system would look like if implemented using Microsoft Foundry and GitHub Copilot's pro-code tooling.

What Gas Town Actually Is

Gas Town is not "Copilot, but bigger."

It is closer to a distributed build system for AI agents.

Its defining characteristics:

Instead of prompting repeatedly, the developer assigns a goal. The system decomposes the goal into sub-tasks, distributes them across workers, supervises output, and merges results.

The novelty is orchestration.

Microsoft's Relevant Building Blocks

To reproduce this architecture in Microsoft's ecosystem, three primary components matter:

1. Microsoft Foundry

Microsoft Foundry (formerly Azure AI Studio) provides:

From a systems perspective, Foundry replaces the need to build a custom multi-agent runtime from scratch.

2. GitHub Copilot SDK & CLI

Copilot's SDK and CLI allow programmatic interaction with code generation models.

This is important because Gas Town's worker agents are essentially programmable coding units.

Using the SDK or gh copilot, it becomes possible to:

That aligns naturally with Gas Town's "short-lived worker" model.

3. Supporting Azure Services

A production implementation would also rely on:

These services provide the operational scaffolding around the agent system.

Gas Town → Microsoft Architecture Mapping

Gas Town Concept What It Does Microsoft Stack Equivalent
MayorPlans work, coordinates agentsFoundry orchestrator agent
Short-Lived Workers (Polecats)Execute focused coding tasksGitHub Copilot SDK sessions
Persistent WorkersMaintain long-running contextFoundry agents with memory
Merge & Review (Refinery)Integrate and enforce qualityPRs + CI + Copilot review
Witness / SupervisionValidate correctnessFoundry evaluation pipelines
Persistent Memory ("Beads")Store durable stateGit + Cosmos DB + Foundry memory
WatchdogsMonitor health and behaviorAzure Monitor + telemetry

This mapping turns Gas Town from a theoretical model into a practical blueprint.

How the System Would Function

Orchestration

In Gas Town, orchestration is handled by ~189k lines of Go.

In a Microsoft implementation, orchestration would be handled by:

The difference is not conceptual — it is infrastructural.

Worker Execution

Each short-lived worker would:

  1. Receive a scoped task
  2. Check out a Git branch
  3. Generate or modify code
  4. Run tests (via CI)
  5. Submit a pull request
  6. Terminate

Persistent workers could maintain architectural context, shared knowledge, or domain understanding across sessions.

State Management

Gas Town's "Beads" system uses Git to maintain persistent work state.

A Microsoft-based version would likely combine:

The pattern remains Git-centric, but memory becomes more flexible.

Supervision and Evaluation

Gas Town includes supervisory roles (Witness, Refinery).

In Microsoft's ecosystem, supervision becomes:

This reduces reliance on manual review while keeping traceability.

Differences Worth Noting

There are meaningful differences between the original Gas Town and a Microsoft implementation.

The tradeoff is flexibility versus operational maturity.

What This Suggests About Multi-Agent Development

Gas Town demonstrates that multi-agent coordination is technically viable.

Microsoft's stack suggests that the same pattern can be implemented using managed services rather than custom infrastructure.

The broader shift is this:

Developers increasingly design systems that coordinate AI agents instead of directly producing code.

Gas Town shows what that looks like in experimental form.

A Microsoft-based implementation shows how it could be structured using enterprise infrastructure.

The architectural pattern remains consistent:

The tools change. The model of development evolves.

Back to All Posts