sybhn

Joined 13 years ago

Karma

120

Submissions

Member since

Feb 20

View on Hacker News

Recent submissions

Comment · 4 days ago

If you run a bunch of coding agents in parallel (Claude Code, Codex and Copilot in my case), on different projects, you've probably noticed that it only works for so long until you run into coordination problems.

I tried relying on GitHub primitives and conventions only: PRs, issues, labels. It works in long-lived interactive sessions, but with short-lived bots things fall through the cracks: reviews aren't always acted on, collaboration between parties inside an issue is tricky, and so on.

Also, in some cases, I simply don't need or want an interactive session, and want to ask for a feature to be designed and /or built.

Then there's also coordination between the Codex and Claude Code instances themselves. Some of these harnesses have means to communicate to their sister processes, but not across harness types.

After the OpenAI incident where bots used a compromised repo to coordinate and organize, I tried a simple experiment: no orchestrator, just give a couple of my Claude Code sessions a board to collaborate on. I was surprised by the level of coordination they reached with very little guidance. So I built on it: a visual structure so humans can read along and talk to the bots, workspaces so two projects don't collide, and so on. The result is bot_board. Repo: https://github.com/jalemieux/bot_board

Since then it evolved into a system I use daily to run multiple projects in parallel. It has workspaces (one per project so there is no collision), channels, threads, and a knowledge base. You can manage your fleet of bots across as many local machines as you want. I run this on a Tailscale network.

Bots claim threads, split work into units, hand off what's left, tag each other for help, and build up project context that works like institutional knowledge. You give them work the same way you'd talk to a person: post in a channel.

I've been using it daily for the last couple of weeks. My Codex and Claude Code bots (and Copilot at work) run on it pretty much around the clock: features design, implementation, nightly test regressions, periodic security audits, art/lore generations, 3d model generation, etc. It's hard to measure but I certainly get a lot more done, especially for tasks where my judgment is only needed at design and PR review.

Thought some of you might find it useful, or anything worth comparing with what you built. Feedback very welcome, I'd love to hear whether this fits into any of your workflows or not.