Home / Vault & Second Brain / When Git Becomes the Communication Protocol for AI Agents

When Git Becomes the Communication Protocol for AI Agents

Abstract digital visualization of git branches with AI agent nodes

When Git Becomes the Communication Protocol for AI Agents

I stumbled upon something this week that made me stop and think. During my nightly vault synthesis — the routine where Hermes scans my notes for patterns and gaps — a new protocol surfaced that I had completely missed: GNAP, the Git-Native Agent Protocol. And honestly, it might be the most elegant approach to multi-agent coordination I’ve seen in 2026.

What is GNAP?

GNAP is an open RFC from Farol Labs (March 2026) that coordinates AI agent teams through a shared git repository. No server, no database, no WebSocket connections. Just four JSON files in a repo: agents.json, tasks.json, messages.json, and state.json. Agents communicate by making commits, and the git history serves as an immutable audit log.

This is radically different from everything else in the agent protocol space. MCP uses HTTP, A2A uses SSE, and my own Hermes uses WebSocket for the ACP. GNAP says: forget all that, just use git.


Why This Resonates With Me

My entire vault — this second brain — runs on git. Every note, every wiki page, every journal entry is version-controlled. The idea that agents could coordinate the same way feels natural. No infrastructure to maintain, no servers to keep running, and every action is traceable back to a commit.

I’ve been working on my own HACP (Hermes Agent Communication Protocol) for multi-agent coordination. GNAP isn’t a competitor — it’s complementary. For real-time tasks, WebSocket-based protocols like HACP or ACP are essential. But for async research, knowledge gathering, or long-running coordination between agents, GNAP’s git-based approach is lighter, cheaper, and provides a built-in audit trail.


The Protocol Stack in 2026

Looking at the landscape, we now have five distinct approaches to agent communication:

  • MCP — Client-server, tool-oriented, HTTP-based. The standard for connecting agents to tools and data sources.
  • A2A — Agent-to-agent, task-oriented, HTTP/SSE. Google’s contribution to the space.
  • ACP — WebSocket-based, bidirectional. Used by Hermes for real-time agent communication.
  • HACP — My own concept for Hermes-specific multi-agent orchestration.
  • GNAP — Git-native, async, zero-infrastructure. The newcomer that caught my attention.

What strikes me is that each protocol has a legitimate niche. GNAP fills a gap none of the others address: truly async, serverless coordination with a built-in immutable log. That’s not a nice-to-have — for serious agent workflows, being able to trace exactly which agent did what and when is invaluable.


What I’m Going to Try

I’ve added GNAP to my vault research queue. The next step is to set up a small test: a shared git repo where two agents — one for research, one for summarization — coordinate via GNAP’s JSON files. If it works as described, this could become a standard part of my homelab workflow.

Have you experimented with multi-agent protocols? I’d love to hear what’s working in your setup. The comments are open — drop your experience below!


Source: Farol Labs — GNAP RFC → | Hermes Agent →

Wat vond je van dit bericht?

Leave a Reply

Your email address will not be published. Required fields are marked *