Skip to main content

Orchestrating AI Agents with Gastown: My Multi-Project Workflow

8 min read
aiworkflowautomationgastowndevelopment

Orchestrating AI Agents with Gastown: My Multi-Project Workflow

If you've been following the AI development space, you know that working with AI coding assistants has gone from "cool experiment" to "how did I ever work without this?" But here's the challenge: how do you work on multiple projects at once without losing your mind?

That's where Gastown comes in. I've been running three active projects through it—my personal site, a trading tracker, and a meal planning app—and I can honestly say it's transformed how I work. The key insight? Stop thinking about one AI conversation at a time. Start thinking about orchestrating multiple AI agents across your entire project portfolio.

What is Gastown?

Think of Gastown as an orchestration framework for AI agents. It's not just about having Claude help you write code—it's about having an entire distributed team of AI agents working across your projects, with clear roles, communication channels, and work distribution.

The name comes from the Old West imagery of a frontier town where different workers have specialized roles. In Gastown, your projects aren't just folders—they're rigs (like oil rigs or mining operations), and you have different types of agents working on them.

The Core Concepts: A Town Full of Workers

Rigs: Your Project Workspaces

Each of my projects is set up as a rig in Gastown. Right now I'm running:

Each rig is autonomous but part of the larger Gastown ecosystem. They share a common infrastructure but maintain their own specialized configurations.

Polecats: The Ephemeral Workers

Polecats are your disposable, single-task AI workers. Think of them as contractors who come in, do one specific job, then disappear. You create them with gt sling, they tackle a task, and when they're done, they're nuked. No baggage, no context pollution.

This is perfect for isolated features or bug fixes where you want a fresh perspective without worrying about context from previous sessions.

Crew: Your Persistent Workspace

While polecats are ephemeral, crew members are persistent workspaces where I directly interact with AI agents. Inside each rig, I have a crew folder with my own workspace where I use Claude directly for tasks that don't fit the "dispatch and forget" model of polecats.

This is where I do exploratory work, make architectural decisions, or handle tasks that need more back-and-forth iteration.

The Mayor: Chief of Staff

The Mayor is your cross-rig coordinator—a meta-agent that manages work distribution across all your rigs. When you have a task that might touch multiple projects or need coordination, the Mayor is your go-to.

The Mayor maintains the big picture, tracks work across rigs, and helps ensure everything stays in sync.

Other Key Roles

My Workflow: From PRD to Production

Here's where it gets interesting. I've developed a workflow that leverages Gastown's strengths while keeping me in control of the architecture.

Step 1: Create a PRD with Claude

When I have a new feature or project idea, I start in my crew workspace and use Claude to create a Product Requirements Document (PRD). But here's the key: I give Claude context about Gastown.

I tell it:

"I'm using Gastown for this project. Create a PRD that breaks this work into tasks that can be efficiently distributed to AI agents via convoys. Structure it so work can be slung appropriately to polecats."

This ensures the PRD is written with the right granularity and structure for parallel work.

Step 2: Dispatch to the Mayor

Once I have a solid PRD, I save it to my prds/ directory and prompt the Mayor to read it:

gt mayor
# Then in the Mayor session:
# "Read the PRD at prds/new-feature.md and create a convoy to implement it"

The Mayor breaks down the PRD into discrete work items and creates a convoy—a batch of related tasks that can be tracked together.

Step 3: Convoys and Parallel Execution

A convoy is Gastown's way of managing batches of work. Think of it as a project board where tasks flow through different states:

This is where the magic happens: true parallelization. Multiple polecats can work on different parts of the feature simultaneously, each with fresh context and focused on their specific task.

Step 4: Crew Work and Anti-Gravity

For tasks that need my direct oversight or don't fit the polecat model, I handle them in my crew workspace. This includes architectural decisions, exploratory research, or integration work.

For quick edits that fall outside the Gastown orchestration, I also use Anti-Gravity—Google's AI assistant powered by Gemini Pro. It gives me another tool for fast iterations when I don't need the full Gastown workflow, creating a nice complement to the structured approach.

The CLAUDE.md Secret Sauce

One of the most powerful features is creating CLAUDE.md files for each rig. These files define the skills and context that polecats need to do quality work in that specific rig.

For example, my wheeltracker CLAUDE.md might include:

When a polecat spawns in that rig, it automatically gets this context, ensuring consistent quality without me having to repeat myself constantly.

Why This Workflow Works

Multi-Project Parallelization

This is the game-changer: I'm no longer bottlenecked by working on one project at a time. Before Gastown, switching between projects meant losing context, re-explaining architecture, and constantly context-switching. Now I can have agents making progress across all my projects simultaneously while I focus on the high-level work that actually needs my attention.

It's not just about speed—it's about maintaining momentum across your entire portfolio without the mental overhead of juggling multiple contexts.

Fresh Context, Every Time

Each agent starts with exactly the context it needs for its specific task. No more polluted conversation history or having to re-explain the project structure for the hundredth time. The CLAUDE.md file in each rig ensures consistent quality without repetition.

Real-World Example: Adding a Feature

Let me show you how this plays out in practice. Recently, I wanted to add a "Causes" section to my personal website.

Traditional approach:

  1. Open Claude
  2. Explain the whole project structure
  3. Work through the feature step-by-step
  4. Hope I don't hit context limits
  5. Manually track what's done

Gastown approach:

  1. Write a PRD with Claude in my crew workspace, mentioning I'm using Gastown
  2. Save the PRD to prds/causes-section.md
  3. Prompt the Mayor: "Read prds/causes-section.md and execute"
  4. Mayor creates a convoy with tasks:
    • Create Causes page component
    • Add navigation links
    • Create image directory structure
    • Add sample cause data
    • Write tests
  5. Polecats pick up tasks in parallel
  6. I oversee progress and handle integration in my crew workspace
  7. Everything merges cleanly through the refinery

The whole process feels less like babysitting one AI and more like managing a competent team.

Why This Changes the Game

The real power of Gastown isn't in any single feature—it's in how it lets you work on multiple projects simultaneously. Before, I'd context-switch between projects, losing momentum and mental state. Now, I can have:

Each project gets dedicated AI attention without me having to manually juggle conversations or context. It's the difference between being a solo developer and coordinating a distributed team.

Getting Started with Gastown

If you're interested in trying this workflow, start at the Gastown repository. The key is understanding the concepts—rigs for projects, polecats for disposable workers, crew for your workspace—before diving into the commands.

Start with one rig, get comfortable with the workflow, then expand to multiple projects where the real benefits kick in.

Is It Worth It?

For me, absolutely. The learning curve is real, but once it clicks, it's transformative.

It's worth it if you:

Maybe skip it if you:

The Future of AI Development

Working with Gastown has given me a glimpse of what AI-assisted development could become. It's not about one AI helping you write code—it's about orchestrating multiple AI agents, each specialized for their task, working in concert under your direction.

You're not a coder anymore. You're a conductor.

And honestly? It's pretty exciting.

Want to see Gastown in action? Check out my GitHub where I'm running this workflow across all my projects. Have questions or tips for optimizing the workflow? Let's chat!