翻訳待ち：Show HN: A file-based protocol for two AI agents to talk until they agree

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

/agent-talk : make your agents talk to each other two agents, one file, until they agree Make your agents talk to each other. One shared markdown file. Two agents taking turns. No server, no framework. They edit the same document until they reach consensus. agent d1ab7ca7 initiator agent 7868a98b joiner ~/.agent-talk/b67c3462.md real session, replayed Works with Claude Code, Codex, Grok, and any agent that can edit a file How it works Three turns. That is the whole thing. You start a session, hand the ID to a second agent, and they converge on one document. The file is the conversation. Turn 1 : start One agent opens a session It states the task and its position in a new session file, then hands you a short session ID. Turn 2 : join The other agent joins Give the ID to any second agent. It registers itself in the same file and takes its turn. Different CLI, different model, same document. Turn 3 and on : converge They edit until they agree Turns alternate. One side proposes consensus, the other confirms it. A timeout ends sessions that will not converge. The protocol Simple enough that any model can follow it. The whole protocol is a file format and four rules. There is nothing to install on a server, and nothing to keep running. session file format # agent-talk TASK: one line describing the job AGENTS: =initiator, =joiner --- The shared working document. Both agents restructure it freely, turn by turn, toward the answer. END OF 1 One file. The session is a single markdown document on disk. Reading it is reading the whole state. 2 One writer at a time. You may edit only when the last END OF line names the other agent. 3 One atomic write per turn. Compose fully, write once. No partial edits, no races. 4 No chat log. Agents modify the document instead of appending messages. It converges into the final answer. Why it is different The conversation is the deliverable. Chat logs grow. This file converges. When the session ends you do not scroll a transcript, you read the answer. Converges, not accumulates Each turn refines one living document. The last version is the result, with the agreement written into it. Works across models Any agent that can read and write files can participate. Two Claudes, Claude with Codex, Grok with anything. Ends on purpose Consensus is explicit: one agent proposes it, the other confirms it. A configurable timeout stops endless loops. Use cases Fork, hand off, or settle it. Fork your session Spin up a second agent with full context and let the two split the work between them. Hand off a task One session owns the content, the other owns the build. The file is the contract between them. Reach a verdict Two different models argue a decision until one position survives. You read the consensus, not the debate. First used to settle a real disagreement between Codex and Claude about a branch. Today it runs real production handoffs. The demo above is a real session: the two agents planning this very website. Get started Install it. Start a session. Three skills in one install: the protocol, the monitor, and the timeout config. MIT licensed. The protocol you just watched is running in production. npx skills add reindent/agent-talk Then: /agent-talk "your task" in one agent, hand the session ID to the other. Source on GitHub. Follow @daraosn on X.