An experiential post about using Gastown multi-agent workspace to demo the Cody Article Writer AI skill, including the BD integration lessons learned along the way.

I spent some time today demoing two AI-powered tools working together: Gastown , a multi-agent workspace manager, and Cody Article Writer , an AI agent skill for structured article writing. The experience surfaced some interesting insights about multi-agent system integration that I want to document.

This demo ran on OpenCode (an interactive CLI tool) using MiniMax M2.1 as the underlying model, not the default Claude. This setup demonstrates how OpenCode can orchestrate agent skills within a multi-agent workspace context while using alternative models.

Diagram showing the Gastown multi-agent workspace orchestrating the Cody Article Writer skill

Gastown convoys track work while the Cody Article Writer skill produces structured articles

What I Was Trying to Do

The goal was straightforward: use Gastown to orchestrate a demo of the Cody Article Writer skill. Specifically:

  1. Create a Gastown convoy to track the demo work
  2. Use an agent to run the Cody Article Writer workflow
  3. Write an article about “The Future of AI-Powered Development”
  4. Track everything through Gastown’s git-backed issue tracker (Beads)
  5. Publish the resulting article to my blog

The demo worked - I ended up with a 4,200-word article about AI development tools. But getting there taught me more than the article itself.

Setting Up the Environment

Gastown organizes work into “rigs” - each rig is typically a git repository. My blog is one such rig, located at ~/gt/my-blog-rig/. The rig has a worktree at refinery/rig/ which contains the actual blog source.

Installing the Cody Article Writer skill was simple. I extracted it into the agent’s skills directory:

bash
unzip cody-article-writer.skill -d .claude/skills/

This made the skill available to any Claude Code instance running in that context. The skill follows the Agent Skills specification and provides a structured workflow: topic ideation → style selection → title/thesis → outline → writing → editorial pass → metadata → export.

The Workflow

Gastown uses “convoys” to track work across multiple issues. I created one for this demo:

bash
gt convoy create "Cody Article Writer Demo" demo-001
gt sling demo-001 my-blog-rig

The sling command was supposed to spawn a “polecat” - an ephemeral worker agent that would pick up the work. This is where things got interesting.

The Integration Issue

The polecat spawned but failed to attach properly. Looking at the logs, I saw errors like:

bd slot set pw-my-blog-rig-polecat-polecat-1 role hq-polecat-role:
  Error: failed to resolve bead hq-polecat-role: no issue found

The naming mismatch was the problem. Gastown uses different prefixes for different contexts:

  • Town-wide beads (hq-* prefix): convoys, deacon, mayor, witness
  • Rig-specific beads (pw-* prefix): project issues for my-blog-rig

The polecat was named with the rig prefix (pw-my-blog-rig-polecat-polecat-1) but trying to access town-wide beads (hq-polecat-role, hq-demo-001). These beads didn’t exist in the expected locations.

Additionally, the BD database had a legacy format issue:

LEGACY DATABASE DETECTED!
This database was created before version 0.17.5 and lacks a repository fingerprint.

The Fix

I resolved this in two steps:

First, migrate the database to the current format:

bash
bd migrate --update-repo-id

This bound the database to the repository and resolved the version mismatch.

Second, create the proper beads with the correct prefixes:

bash
bd new "Write demo article using cody-article-writer skill" --id hq-demo-001
gt convoy add hq-cv-lg5jg hq-demo-001

Once the beads existed with the right names, the convoy tracking worked correctly. When I closed hq-demo-001, the convoy auto-closed, showing 1/1 completed.

What Cody Article Writer Produced

Despite the integration hiccups, the skill itself worked beautifully. I gave it a topic (“The Future of AI-Powered Development”) and a style guide (“Tech Professional” - tone 7/10, technical 8/10). The skill guided me through:

  1. Topic refinement - Clarifying the thesis
  2. Style selection - Choosing a professional voice
  3. Title and thesis - “AI-assisted development is not replacing developers but amplifying their capabilities”
  4. Outline generation - Five sections covering current state, capabilities, human-AI partnership, team implications, and future preparation
  5. Article writing - Full draft with sections, examples, and formatting
  6. Editorial pass - Optional polish (I skipped this for the demo)
  7. Metadata - Title, description, keywords for frontmatter
  8. Export - Markdown file with YAML frontmatter

The output went to cody-projects/article-writer/articles/the-future-of-ai-powered-development.md, and has since been published to the blog:

The Future of AI-Powered Development

Key Takeaways

  1. Multi-agent systems need consistent naming conventions. When work spans town-wide and rig-specific contexts, bead prefixes must align.

  2. Legacy database migrations matter. The BD database had been created before version 0.17.5. Running migrations proactively prevents integration issues.

  3. Convoys provide good visibility. Even when the polecat failed, the convoy system tracked the work and I could see what was supposed to happen.

  4. Agent skills integrate cleanly when configured correctly. Cody Article Writer is designed to be dropped into any Claude Code environment. Once the path was set up (cody-projects/article-writer/), it just worked.

  5. State persistence is valuable. The skill saved drafts, styles, and exports. If I’d been interrupted, I could resume with “continue my article”.

The Meta-Lesson

What’s interesting is that this post itself is meta - I’m writing about the experience of writing an article. The gastown + cody integration creates a recursive loop: agents using tools that help agents.

The real value isn’t any single article. It’s the pattern:

  • Gastown provides orchestration and state tracking
  • Agent skills provide specialized capabilities
  • Beads provides git-backed persistence
  • The combination creates a workspace where AI agents can collaborate on complex, multi-step tasks

I’m planning to use this setup for future writing projects. The structure forces me to think through topics before diving in, and the git-backed state means I can work across sessions without losing context.

Tags: Gastown, Cody Article Writer, AI Agents, Multi-Agent Systems, Beads

Join the Conversation

Comments section coming soon...