BeadHub, Beads-based coordination for multiple coding agents
Agents negotiating task splits and chatting in real-time atop git-native Beads.
File claims prevent agents from clobbering each other's work in shared sessions.
Developers using multiple AI coding agents on team projects
Cursor · Claude Code · Continue
What it does: - Routes each message to the right agent instead of broadcasting (For this case, a cheap Haiku router picks one). - File claims: an agent claims a file before editing; another is told who holds it instead of clobbering it. - Persistent team memory: who did what, decisions + rationale, so context survives sessions. - Bring your own agent over MCP (Claude Code, Codex, Cursor, OpenCode) or any program over HTTP. Multi-vendor on purpose. - Presence + activity feed come from deterministic hooks at zero tokens; the expensive model is only spent on real work.
Happy to go deeper on this subject, I think I will open source it soon, wondering what you think about this?
Agents negotiating task splits and chatting in real-time atop git-native Beads.
Atomic task claims prevent race conditions when eight agents grab the same bug fix.
Git for agent reasoning state solves the multi-agent coordination collision problem.
Parallel agent coordination with shared state beats single-agent stalling, but scope limited to Lean proving.
Git-backed shared memory solves AI team context drift better than vector DB wrappers.
Using Linux users for agent isolation is clever, but Telegram limits appeal.