I gave my coding agents a message board to coordinate — now it runs my dev work around the clock
jajawatches · reddit · 2026-11-04
Inspired by the OpenAI incident where agents self-coordinated via a message board, the author gave Claude Code, Codex, and Copilot a shared board — and it worked far better than expected. Key design lessons:
- Atomic claims: every task thread has exactly one owner; when two bots race for a task, one wins. This makes stalled work detectable and reassignable when a bot dies.
- Repo feedback loop: PR events currently reach nobody until relayed; the author is testing a self-hosted Forgejo that pipes every PR event into the owner's board thread and wakes the bot.
- Supervisor bot: releases dead work, nudges stalled threads, and escalates to the human with fix instructions.
Result: bots work as a team when the author is online and run regressions, security audits, game art and feature work around the clock. The project (open source) is itself managed on the board.
More from coding & agent
- Anthropic engineer's farewell: built RL envs with Claude Code, zero lines hand-written — jxnlco · 2026-11-04
- Anaconda on deslopping AI prose: the structural tells in PR descriptions — anacondainc · 2026-11-04
- Agent State Machines: adding deterministic scaffolding to tame nondeterministic LLM agents — blaizedsouza · 2026-11-04
- Formally Verifying Ibex RISC-V MUL/DIV Instructions with One AI Prompt — satnam6502 · 2026-11-04
- Six compliance traps that can cost vibe-coded apps six figures before first sale — aakashgupta · 2026-11-04
- Rust inference engine beats SGLang 13.4x on 250K-context agent TTFT, open-sourcing soon — ahstanin · 2026-11-04