GNAP: git-native task coordination for AutoChain's multi-agent evaluation workflows

[ori-cofounder]

Proposal: GNAP as a coordination layer for AutoChain's automated agent evaluation

AutoChain is a lightweight generative agent framework focused on rapid iteration — easy customization and automated multi-turn evaluation via simulated conversations. As you build more complex evaluation pipelines (multiple agents being evaluated in parallel, coordinator + specialist patterns), coordination becomes important.

GNAP (Git-Native Agent Protocol) provides a minimal coordination substrate: a git repo with board/todo/ → board/doing/ → board/done/. No extra dependencies — AutoChain is already lightweight and GNAP keeps it that way.

Applied to AutoChain's workflow evaluation:

AutoChain evaluates agents with simulated conversations. GNAP could coordinate parallel evaluation runs:

board/todo/eval-agent-v1-scenario-happy-path.md    ← Evaluator creates
board/todo/eval-agent-v1-scenario-edge-case.md
board/todo/eval-agent-v1-scenario-error-recovery.md

board/doing/eval-agent-v1-scenario-happy-path.md   ← AutoChain eval agent claims
board/doing/eval-agent-v1-scenario-edge-case.md     ← Another eval agent claims
board/done/eval-agent-v1-scenario-happy-path.md     ← Eval results + success rate

Parallel evaluation without conflicts, persistent results across restarts, and a full audit trail of which scenarios passed/failed — all via git. No additional infrastructure.

GNAP is particularly well-suited for AutoChain's OpenAI function-calling agents since those agents already structure their outputs — GNAP task files can carry structured specs that map directly to function call parameters.

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