Our Claude Code agents now message each other before two of them edit the same file
retained text
My team runs a lot of Claude Code sessions in parallel, a few per person, each in its own worktree or on its own laptop. Each of those agents works blind. It doesn't know what the other agents, mine or my teammates', are working on, what they've decided, or what they've already figured out. So one changes the shape of an API response while another builds against the old one, two of them dig into the same bug, or one works out from scratch something a colleague's agent already knew. Conflicting PRs are just the moment you notice. So we built murmur, to keep the agents in touch so they work more like one team. A small daemon on each laptop reads what Claude Code and git already record (sessions, their titles, branches, changed files), so the agents don't fill anything in. Each session gets three MCP tools: who shows what every other agent is working on, send messages an agent or a person, and btw asks another Claude Code session a question without interrupting it (a throwaway copy of its conversation answers). An agent about to change an API can ask the agent building the client what it relies on, and an agent that's stuck can ask the one that already worked it out. There's one moment where we make sure they talk: a PreToolUse hook. When an agent is about to edit a file another checkout is changing, the edit is turned back once with the other agent's name. In practice the agent then messages the other one, they agree on who does what, and the edit goes through. That's what the video shows, but it's the safety net, not the point. The video is a one-minute explainer we made. The README has an unedited recording of two sessions doing exactly this. How Claude helped: almost all of murmur's code was written by Claude Code, with several sessions working at once on different parts (the laptop daemon, the hub, the terminal console) while we reviewed and steered. That's how we ran into the problem in the first place, since our own sessions kept stepping on each other. As soon as an early version worked, those sessions ran on it, and watching them shaped the design. In two days they sent each other 147 messages, 48 of them between different people's agents. That first version warned about every overlap and was noisy (9 in 10 warnings were repeats), which is why it now speaks up once, right when an agent is about to edit. It's free to try on your own: brew install eledoneai/tap/murmur, then murmur solo, and start two sessions in two worktrees of the same repo. For a team you host one small hub yourselves. There's no central service, and code, prompts and transcripts stay on the laptops. It also works with Codex, Cursor, OpenCode, Copilot CLI and Gemini CLI, so not everyone has to use the same tool. It's Apache-2.0 and there's nothing to sell. It's also very new, so I'd like to hear where it breaks for you, and how you keep several agents on the same page today. [https://github.com/EledoneAi/murmur](https://github.com/EledoneAi/murmur) Edit: rewritten for clarity. The first version led with merge conflicts and made murmur sound like a conflict tool, which it isn't. https://reddit.com/link/1xjjrc2/video/f0hlhcggtmyh1/player
document fetched 2026-10-31 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1xjjrc2/our_claude_code_agents_now_message_each_other