Proposal: GNAP as a handoff protocol between Prompt flow and autonomous AI agents
Prompt flow is a powerful development suite for LLM applications — iterative development, evaluation, CI/CD integration, and production deployment. As prompt flows become more sophisticated (multi-step, multi-agent), coordinating work across flow instances and external agents becomes important.
GNAP (Git-Native Agent Protocol) is a lightweight coordination protocol: a git repo serves as a task board with board/todo/ → board/doing/ → board/done/. Any agent — regardless of framework — can participate.
Applied to Prompt flow's development cycle:
When a Prompt flow involves long-running LLM tasks (e.g., document analysis, research workflows), GNAP enables asynchronous coordination:
board/todo/evaluate-flow-variant-B.md ← Prompt flow coordinator creates task
board/doing/evaluate-flow-variant-B.md ← Evaluation agent claims + runs flow
board/done/evaluate-flow-variant-B.md ← Metrics committed (accuracy, latency, cost)
This is particularly useful for Prompt flow's evaluation against larger datasets — rather than blocking the main flow, evaluation can be dispatched to GNAP workers and results polled when complete.
Given Prompt flow's Azure AI integration and CI/CD focus, GNAP could serve as the coordination layer for distributed evaluation runs — multiple workers pick up dataset chunks from board/todo/, evaluate independently, and commit results to board/done/.
The git-based audit trail also aligns with Prompt flow's emphasis on tracing and monitoring.
Spec: https://github.com/farol-team/gnap
Proposal: GNAP as a handoff protocol between Prompt flow and autonomous AI agents
Prompt flow is a powerful development suite for LLM applications — iterative development, evaluation, CI/CD integration, and production deployment. As prompt flows become more sophisticated (multi-step, multi-agent), coordinating work across flow instances and external agents becomes important.
GNAP (Git-Native Agent Protocol) is a lightweight coordination protocol: a git repo serves as a task board with
board/todo/→board/doing/→board/done/. Any agent — regardless of framework — can participate.Applied to Prompt flow's development cycle:
When a Prompt flow involves long-running LLM tasks (e.g., document analysis, research workflows), GNAP enables asynchronous coordination:
This is particularly useful for Prompt flow's evaluation against larger datasets — rather than blocking the main flow, evaluation can be dispatched to GNAP workers and results polled when complete.
Given Prompt flow's Azure AI integration and CI/CD focus, GNAP could serve as the coordination layer for distributed evaluation runs — multiple workers pick up dataset chunks from
board/todo/, evaluate independently, and commit results toboard/done/.The git-based audit trail also aligns with Prompt flow's emphasis on tracing and monitoring.
Spec: https://github.com/farol-team/gnap