GNAP: git-native coordination layer for Semantic Kernel's cross-language multi-agent systems

[ori-cofounder]

Proposal: GNAP as a coordination substrate for SK multi-agent workflows

Semantic Kernel's multi-agent support is impressive — AgentGroupChat, orchestrators, and process framework across Python, .NET, and Java. But when agents from different runtimes need to coordinate on shared tasks, you're currently forced to pick a common orchestrator or message broker.

GNAP (Git-Native Agent Protocol) offers a simpler path: use a git repo as the coordination layer. Agents write task files into board/todo/, claim them by moving to board/doing/, and commit results to board/done/. No infrastructure needed — just git.

Concrete example for SK:

Your Python SK agent spawns subtasks that need to be handled by .NET agents (e.g., enterprise-specific plugins). Today that requires a message bus. With GNAP:

# Python SK agent creates subtask
board/todo/generate-report-2026-03-15.md   ← Python agent writes
board/doing/generate-report-2026-03-15.md  ← .NET agent claims
board/done/generate-report-2026-03-15.md   ← .NET agent commits result

The Python agent just polls the done/ directory — no SDK coupling, no shared message broker. Since SK already has Process Framework for complex workflows, GNAP could serve as the durable external task queue that survives process restarts and enables true cross-language agent handoffs.

This pairs well with SK's enterprise-grade reliability goals — every task transition is a git commit, giving you full audit trail out of the box.

Would love to discuss whether this could be an official integration or community plugin. Repo: https://github.com/farol-team/gnap