GNAP: git-native coordination protocol for AgentVerse's multi-agent simulation environments

[ori-cofounder]

Proposal: GNAP as a task coordination layer for AgentVerse's multi-LLM environments

AgentVerse provides a multi-agent simulation framework for both task-solving and simulation scenarios. Your task-solving mode organizes agents into discussion groups that iteratively solve complex problems — coordination between agents is central to the design.

GNAP (Git-Native Agent Protocol) could enhance AgentVerse's inter-agent coordination: a git repo as a persistent task board with board/todo/ → board/doing/ → board/done/. Tasks are files, transitions are commits, no extra infrastructure.

Applied to AgentVerse's task-solving pipeline:

AgentVerse's discussion-based task solving (recruiting agents, discussing, executing) maps naturally to GNAP:

board/todo/subtask-implement-sort-algo.md   ← Discussion phase decomposes task
board/todo/subtask-write-unit-tests.md
board/doing/subtask-implement-sort-algo.md  ← Coder agent claims
board/doing/subtask-write-unit-tests.md     ← Tester agent claims
board/done/subtask-implement-sort-algo.md   ← Implementation committed
board/done/subtask-write-unit-tests.md      ← Tests committed

For AgentVerse's simulation mode, GNAP provides an auditable log of all agent interactions — a valuable research artifact for studying emergent coordination behavior in multi-agent systems.

Given AgentVerse's research orientation (arxiv:2308.10848), GNAP's audit trail could support reproducible experiments — you can replay any multi-agent coordination scenario from the git history.

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