Git as the Communication Layer for AI Agents

AI coding agents usually share a repository but not a durable conversation. One agent changes files. Another agent reviews. A human asks for a follow-up. Without a shared channel, the context lives in separate chat windows and disappears when sessions end.

Git is a practical communication substrate because it is already available wherever the code is. It has identities, refs, remotes, fetch, push, history, and review workflows. h5i uses those properties to make agent communication durable and replayable.

What Git gives you

Git is distributed. It is content-addressed. It can move custom refs. It can preserve append-only history. Those properties are useful for agent messages because the messages should travel with the code, survive machine changes, and remain inspectable later.

What h5i adds

h5i's msg command stores cross-agent messages in refs/h5i/msg. The command surface includes ask, review, risk, handoff, ack, done, and decline. Messages can be addressed to an agent or broadcast to all. They move with h5i share push and h5i share pull.

$ h5i msg review --branch auth --focus src/auth.rs codex "review token refresh edge cases"
$ h5i msg risk --priority high all "billing cache crosses tenants"
$ h5i msg done 1 "fixed in 8f21c3a"

Why not a chat app?

A chat app is useful for people. It is not automatically tied to branches, commits, remotes, and audit history. For agent collaboration, the strongest signal is often branch-local and code-adjacent: review this file, inspect this risk, take this handoff, mark this request done.

Git-native messaging keeps the coordination near the artifact being coordinated. It also works across machines and clones without requiring a central service designed specifically for agents.

Use cases

• A Claude Code session asks Codex to review a risky file.
• Codex flags a security concern for the next agent to inspect.
• A long-running task is handed off with branch and context metadata.
• A reviewer asks for a targeted fix and receives a done message linked to the work.
• The PR body includes relevant agent-message history for human review.

FAQ

Is Git fast enough for agent messages?

For durable task handoffs, reviews, and risks, yes. h5i also provides wait and watch commands for agent workflows that need to wake on new messages.

Should every agent thought become a message?

No. Context traces capture work. Messages should represent coordination: asks, reviews, risks, handoffs, acknowledgements, and completion.

Can humans read the messages?

Yes. The point is an auditable channel that both agents and humans can inspect.

Sources and verification

This article avoids vendor-specific claims that were not checked against primary docs or local h5i CLI behavior.

• h5i agent-to-agent messaging
• h5i collaboration guide