GNAP: git-native coordination for AgentRxiv's multi-agent collaborative research #113
Description
Proposal: GNAP as a coordination protocol for AgentRxiv's inter-agent research workflows
Agent Laboratory with AgentRxiv is doing something genuinely exciting — autonomous research agents that upload, retrieve, and build on each other's work to make cumulative progress. This is exactly the use case GNAP is designed for.
GNAP (Git-Native Agent Protocol) coordinates multiple AI agents via a shared git repo: board/todo/ → board/doing/ → board/done/. It's a natural complement to AgentRxiv — while AgentRxiv handles the research artifact layer (papers, results), GNAP handles the coordination layer (task assignment, work in progress, completion).
Applied to AgentLaboratory's three-phase workflow:
board/todo/lit-review-quantum-ml-2026-03.md ← Literature Review agent creates
board/doing/lit-review-quantum-ml-2026-03.md ← Agent claims + queries arXiv
board/done/lit-review-quantum-ml-2026-03.md ← AgentRxiv preprint path committed
board/todo/experiment-quantum-ml-hypothesis.md ← Experiment phase agent picks up
board/doing/experiment-quantum-ml-hypothesis.md ← Agent claims, runs HuggingFace
board/done/experiment-quantum-ml-hypothesis.md ← Results + code committed
board/todo/write-report-quantum-ml.md ← Report agent synthesizes
GNAP prevents duplicate work (atomic task claiming), provides a full audit trail of the research pipeline, and enables parallel execution of independent experiments. Since Agent Laboratory agents already write files and run code, GNAP participation requires no new capabilities.
The git history of a GNAP-coordinated AgentRxiv research run would be a fascinating dataset — showing exactly how collaborative AI research unfolds at the task level.