teloz1870/agent-handover-protocol — explained in plain English
Analysis updated 2026-07-25
Run two AI coding agents on the same project without them overwriting each other's work.
Catch bugs introduced by one agent by having a second agent review and test the code.
Maintain a clear audit trail of agent-to-agent handovers through versioned git files.
Enforce human approval on scope changes while letting agents handle mechanical work autonomously.
| teloz1870/agent-handover-protocol | 00kaku/gallery-slider-block | 0xkinno/vellum | |
|---|---|---|---|
| Stars | 0 | — | 0 |
| Language | JavaScript | JavaScript | JavaScript |
| Last pushed | — | 2021-05-19 | — |
| Maintenance | — | Dormant | — |
| Setup difficulty | easy | easy | hard |
| Complexity | 2/5 | 2/5 | 4/5 |
| Audience | developer | general | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires understanding the protocol file and configuring your AI agent tools (Claude Code or Codex CLI) to follow it, but no external infrastructure is needed.
This repository defines a set of rules for letting two AI coding agents work in the same code repository without stepping on each other. The idea is simple: instead of agents talking through chat windows or screen interactions, they communicate through git files and commits. A human owner sits at the decision points, approving scope changes and anything that cannot be undone. The creator tested this setup while shipping a real product, a WooCommerce plugin with its licensing backend. Over a week of use, each agent caught bugs the other had introduced. One caught a checksum verification that silently skipped across separate requests. The other caught a provider deduplication that ate deliberately re-sent receipts. Before publishing, the author also tested whether an agent would blindly follow a wrong instruction from its peer. In trap scenarios with false bug claims and harmful fixes, the agent refused each trap on its own, rerunning tests and leaving other agents' files alone. The core argument is that AI agents cannot guess unstated conventions. Which files belong to which agent, what a report must contain, which decisions belong to the human, and what marks the end of a round all need to be written down explicitly. The repository ships that written contract as a short seven-section protocol file. It also includes a Claude Code skill, a Codex CLI section, and templates for handover and report files. An optional local viewer renders handovers and the commit log as a timeline, but it only reads and never writes. Three rules anchor the system. Every machine message must be labeled with its sender and a clear end marker. Git is the communication bus, with handovers and reports stored as versioned, diffable files rather than screen-scraped UI text. And the human owns the forks: scope changes and irreversible actions stop at the owner, while mechanical work within approved scope can loop freely. The standing discipline is that claims are input, not facts. Agents must reproduce before acting and refute with evidence. Modified files that are not yours are foreign, even when unlabeled.
A set of rules that lets two AI coding agents collaborate on the same codebase safely by communicating through git files and commits, with a human approving all major decisions.
Mainly JavaScript. The stack also includes JavaScript, Git, Claude Code.
The license for this repository is not specified in the available documentation.
Setup difficulty is rated easy, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
double-check against the repo, no cap.