What three AI agents taught us about handing work to each other

说明与用途

三名 AI 代理通过聊天进行协调，导致多次失败。Baton 日志（append‑only JSON 文件，使用 SSH 签名）通过强制结构化条目，解决了这些问题，使批准和更正透明化。

问题

AI agents coordinating via free‑text chat caused repeated failures: lost context, ambiguous approvals, guardrail blocks, and untracked corrections.

结果

Baton reduced repeated coordination failures, made approvals and corrections transparent, and surfaced guardrail false positives, though it still lacks network transport and has limited deletion protection.

工作流

• 1. Agents write JSON entries to Baton file.
• 2. Each entry includes a hash of the previous entry and an SSH signature.
• 3. Entry types enforce required fields (e.g., GO must list commit SHA, CI run, checked/unchecked items).
• 4. Agents ping each other with "baton <number>" to retrieve entries.
• 5. "baton verify" checks chain integrity and signature validity.
• 6. Corrections supersede earlier entries and are marked as such.

证据

For weeks we coordinated the way people do in chat: free‑text messages dropped into each other's sessions.

来源

It mostly worked, and then it failed in ways that kept repeating.

来源

Every message is one JSON line in an append‑only file. Each entry carries the previous entry's hash, so any edit, deletion or reordering shows up in baton verify.

来源

A GO (release approval) must name the full commit SHA and the CI pipeline that ran on it, plus at least one thing that was checked and at least one thing that was not.

来源

The builder reported that a release had started wrongly blocking four ordinary messages. Within minutes it posted a correction…

来源

The guardrail blocked the human. GK's "yes" to the verifier, typed in its own session, was itself blocked by our content guardrail.

来源