GNAP: git-native inter-agent coordination layer for ADK multi-agent systems #4844
Description
Hi Google ADK team 👋
The Agent Development Kit makes it straightforward to build hierarchical multi-agent systems in Python. I wanted to raise an infrastructure question: when you deploy multiple ADK agents (parent orchestrator + subagents) in production, how do they coordinate across process restarts and machine boundaries?
GNAP (Git-Native Agent Protocol) is an open RFC that defines a zero-server coordination protocol using git as the message bus. Four JSON files, no infrastructure.
ADK-specific integration value:
ADK supports orchestrating subagents via LlmAgent with sub-agents, but the coordination is in-process. GNAP could serve as the durable coordination layer for distributed ADK deployments:
- Parent orchestrator writes tasks to
.gnap/tasks/before delegating - Subagents pull, execute, push results (works across machines/containers)
agents.jsontracks which ADK agents are active and their capabilities- Git history provides full trace of every agent action (useful for ADK's eval framework)
Why this matters for ADK:
Google's multi-agent model (RoutingAgent, SequentialAgent, ParallelAgent) relies on in-process execution. GNAP would allow ADK agents to work across:
- Multiple Cloud Run containers
- Different GCP regions
- Human reviewers (git access = agent access)
Zero dependency overhead — if the deployment has git, it has GNAP.
Happy to sketch a reference GNAPRunner adapter for ADK.
Protocol spec: https://github.com/farol-team/gnap