GNAP: git-native task board for XAgent's inner/outer loop multi-agent coordination

Proposal: GNAP as a coordination layer for XAgent's PlanAgent + ToolAgent architecture

XAgent's dual-loop architecture (PlanAgent for task decomposition + ToolAgent for execution) is a solid foundation for autonomous task solving. When scaling to multiple XAgent instances working in parallel on decomposed subtasks, you need a coordination protocol.

GNAP (Git-Native Agent Protocol) provides this: a git repo as a shared task board with board/todo/ → board/doing/ → board/done/. XAgent's PlanAgent can write subtasks to board/todo/, and multiple ToolAgent instances can claim and execute them in parallel.

Applied to XAgent's task decomposition:

XAgent already breaks complex tasks into subtasks. GNAP makes this coordination explicit and persistent:

board/todo/subtask-1-search-papers.md      ← PlanAgent decomposes
board/todo/subtask-2-analyze-results.md
board/doing/subtask-1-search-papers.md     ← ToolAgent instance 1 claims
board/done/subtask-1-search-papers.md      ← Results + tool logs committed

board/doing/subtask-2-analyze-results.md   ← ToolAgent 2 picks up after subtask 1

XAgent runs in Docker containers — which already have git available. GNAP adds zero infrastructure and the coordination board persists across container restarts. The git history also provides an auditable record of the entire task decomposition and execution trace.

GNAP's simplicity aligns with XAgent's design goal of being general-purpose and extensible — any new tool or agent type can participate in GNAP without SDK changes.

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

Start by reading XAgent's PlanAgent and ToolAgent architecture, then review the linked GNAP specification and its board/todo, board/doing, and board/done workflow. Define the integration boundaries and acceptance criteria for persistent multi-agent coordination before identifying implementation files or tests.