GNAP: git-native coordination for fast-agent multi-model pipelines

[ori-cofounder]

Hi fast-agent team 👋

fast-agent's composition model (chains, parallel agents, routers, evaluator-optimizers) is one of the cleanest approaches to multi-agent orchestration I've seen. I wanted to raise a question about cross-process and multi-machine coordination.

The gap: fast-agent's workflow composition is elegant but in-process. If you're running a parallel of 5 agents and want to distribute them across machines, or resume after a failure, there's no built-in mechanism.

GNAP is an open RFC for zero-server agent coordination. The protocol: agents read tasks/*.json, do work, write results, commit and push. Git is the only infrastructure needed.

fast-agent integration idea:

fast-agent's @fast.chain() and @fast.parallel() decorators could optionally back their task queues with GNAP:

@fast.chain(
    name="research_pipeline",
    backend="gnap",  # durability via git
    gnap_repo="./agent-workspace"
)
async def research_chain():
    ...

This would make fast-agent workflows:

• Resumable — crashed mid-chain? Pull and continue
• Distributable — run chain steps on different machines
• Auditable — full git history of every agent step
• Interoperable — any MCP-compatible agent can participate via GNAP

Given fast-agent's focus on building with open-source models, the zero-infrastructure aspect of GNAP seems especially aligned.

Happy to discuss or prototype a GNAPBackend for fast-agent.

Protocol spec: https://github.com/farol-team/gnap

[iqdoctor]

Try this: https://github.com/steveyegge/beads

I'm using version 0.49.6 (the latest stable version based on SQLite).