GNAP: git-native task coordination to complement Agent-Git versioning #5
Description
Hi — Agent-Git's approach to versioning agentic state with git is exactly the right instinct. You're solving "how do we version-control an agent's decisions over time."
We're building GNAP which tackles a related but different layer: how do multiple agents coordinate on shared tasks, also using git as the substrate.
Where they complement each other:
- Agent-Git = version control for individual agent memory/state (branching, rollback, RL)
- GNAP = coordination protocol between agents (task claiming, work queues, audit trail)
In practice: a LangGraph agent using Agent-Git for its own state could participate in a GNAP coordination board for multi-agent task distribution. The git repo becomes both the agent's memory and the coordination medium.
The MDP/RL angle in Agent-Git is interesting — has anyone explored using git commit history as the reward signal for RL? That would be a natural fit since GNAP commits have structured outcomes in the task files.