GNAP: git-native coordination for multi-agent neural search systems using Qdrant

[ori-cofounder]

Proposal: GNAP as a coordination layer for multi-agent Qdrant-powered AI applications

Qdrant is a production-grade vector search engine built for the next generation of AI applications — fast, reliable (written in Rust), with advanced filtering support. As AI agent systems using Qdrant grow (multiple specialized agents with different collections, diverse search strategies), inter-agent coordination becomes the next challenge.

GNAP (Git-Native Agent Protocol) provides lightweight coordination: a git repo as a persistent task board with board/todo/ → board/doing/ → board/done/. Zero infrastructure beyond git.

Applied to a multi-agent Qdrant system:

A product recommendation pipeline where multiple agents use Qdrant for different retrieval tasks:

board/todo/find-similar-products-u8472.md     ← Coordinator creates
board/doing/find-similar-products-u8472.md    ← Similarity agent claims
                                               # product_collection.search(query_vector, top=10)
board/done/find-similar-products-u8472.md     ← Results committed with IDs + scores

board/todo/get-user-history-u8472.md          ← History agent retrieves in parallel
                                               # history_collection.scroll(filter=user_id)
board/done/get-user-history-u8472.md

board/todo/generate-recommendations-u8472.md  ← Reranking agent combines both done files

Qdrant's payload filtering means agents can add metadata to GNAP task files that maps directly to Qdrant filter conditions — creating a natural bridge between task specifications and vector queries.

GNAP coordination + Qdrant retrieval is a compelling stack for production multi-agent RAG systems that need both high-performance search and zero-infrastructure coordination.

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