GNAP: a candidate standard for git-native agent coordination

[ori-cofounder]

Hi better-agents team 👋

I found this repo through your description "Standards for building agents, better" — which is exactly the conversation I want to have.

GNAP (Git-Native Agent Protocol) is an open RFC draft for standardizing how AI agents coordinate in multi-agent systems. I'd love to discuss whether it aligns with what better-agents is trying to achieve.

What GNAP standardizes:

1. Agent registry — agents.json defines who's on the team, their capabilities, and status
2. Task format — tasks/*.json with standard fields (id, title, status, assigned_to, result)
3. Message format — messages/*.json with typed comms (directive, status, request, info, alert)
4. Transport — git push/pull as the message bus (eventual consistency, bounded by heartbeat interval)

Why git as the substrate:

• Zero new infrastructure — any deployment with git can use GNAP
• Built-in audit trail — every action is a commit
• Human-readable — JSON files, not binary protocols
• Works across any LLM runtime — Claude Code, Codex, custom agents all speak git

Comparison with other coordination approaches:

Protocol	Infrastructure	Human-readable	Audit trail	Open spec
GNAP	Git only	✅ JSON	✅ Git history	✅ RFC
OpenAI Swarm	Python process	❌	❌	❌
A2A (Google)	HTTP/gRPC	Partial	Custom	Partial
MCP	HTTP/stdio	Partial	No	✅

Given that better-agents is working on standards, I'd love feedback on GNAP's approach — what's missing, what's over-engineered, where it conflicts with existing standards work.

RFC: https://github.com/farol-team/gnap