Proposal: GNAP as a coordination layer for MindSearch's concurrent WebSearcher agents
MindSearch's mimicking of human mind search — a WebPlanner that breaks queries into parallel WebSearcher subtasks — is elegant. The concurrent multi-query search (improved in v0.5 with Lagent) is the key performance driver. Coordinating these parallel searchers is exactly what GNAP is designed for.
GNAP (Git-Native Agent Protocol) provides durable coordination: board/todo/ → board/doing/ → board/done/. Each WebSearcher claims a search task, completes it, and commits results.
Applied to MindSearch's parallel search architecture:
When WebPlanner decomposes a complex query into parallel sub-queries:
board/todo/search-quantum-error-correction-basics.md ← WebPlanner creates
board/todo/search-quantum-error-correction-recent.md
board/todo/search-quantum-hardware-implementations.md
board/doing/search-quantum-error-correction-basics.md ← WebSearcher 1 claims
board/doing/search-quantum-error-correction-recent.md ← WebSearcher 2 claims in parallel
board/done/search-quantum-error-correction-basics.md ← Results committed
board/todo/synthesize-quantum-overview.md ← WebPlanner picks up all done results
This provides crash-safe parallel search — if a WebSearcher crashes mid-search, another can reclaim the task from board/todo/. Given MindSearch's deployment on Puyu (production scale), this durability is valuable.
GNAP also captures the full search decomposition history in git — a useful research artifact for studying how the WebPlanner decomposes different query types.
Spec: https://github.com/farol-team/gnap
Proposal: GNAP as a coordination layer for MindSearch's concurrent WebSearcher agents
MindSearch's mimicking of human mind search — a WebPlanner that breaks queries into parallel WebSearcher subtasks — is elegant. The concurrent multi-query search (improved in v0.5 with Lagent) is the key performance driver. Coordinating these parallel searchers is exactly what GNAP is designed for.
GNAP (Git-Native Agent Protocol) provides durable coordination:
board/todo/→board/doing/→board/done/. Each WebSearcher claims a search task, completes it, and commits results.Applied to MindSearch's parallel search architecture:
When WebPlanner decomposes a complex query into parallel sub-queries:
This provides crash-safe parallel search — if a WebSearcher crashes mid-search, another can reclaim the task from
board/todo/. Given MindSearch's deployment on Puyu (production scale), this durability is valuable.GNAP also captures the full search decomposition history in git — a useful research artifact for studying how the WebPlanner decomposes different query types.
Spec: https://github.com/farol-team/gnap