GNAP: git-native coordination that AgentOps can observe and trace end-to-end

[ori-cofounder]

Proposal: GNAP as a coordination protocol with first-class AgentOps observability

AgentOps provides observability and DevTools for AI agents — session tracking, cost monitoring, replay, and now an MCP server. The core challenge you solve is making agent behavior visible and debuggable. GNAP adds a layer that's directly traceable.

GNAP (Git-Native Agent Protocol) uses git as the coordination substrate: tasks move through board/todo/ → board/doing/ → board/done/ via commits. Every task transition is a git commit — with timestamp, agent identity, and content.

Why GNAP + AgentOps is a natural fit:

1. Git commits as observable events: Each GNAP state transition (todo→doing, doing→done) maps perfectly to an AgentOps event — agent ID, timestamp, task content
2. Multi-agent session reconstruction: From the git log, AgentOps can reconstruct the full coordination graph across all agents
3. Cost attribution: Which agent spent what on which GNAP task — combining AgentOps LLM cost tracking with GNAP task metadata

# GNAP task transition becomes AgentOps event
agentops.record(ActionEvent(
    action_type="gnap.task.claimed",
    params={"task": "board/doing/analyze-dataset-47.md", "agent": agent_id},
    returns={"status": "in_progress"}
))

GNAP could integrate with AgentOps' existing framework integrations (LangChain, AutoGen, CrewAI) — any agent using those frameworks that participates in a GNAP board would automatically have its coordination events traced.

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