I extended Karpathy’s “Agentic Engineering” using Google’s Agents CLI so that a Retrieval‑Augmented Generation (RAG) agent can pause after drafting a reply, wait for human approval, and later resume without re‑running retrieval or repeating side‑effects.
The retrieval count remains one, and the answer is published once.
The agent retrieves four policy passages, writes a cited customer reply, asks for approval, and exits.
Submitting the approval again returns already_published instead of calling the publication tool twice.
Replaying the prompt is the wrong recovery mechanism here since it runs retrieval again, may produce a different answer, and can repeat the external action.
Google's Agents CLI lets a coding agent build this directly. It installs seven focused skills that cover ADK development, testing, and deployment to Agent Runtime.
The app retrieves the evidence, drafts the reply, and exits with a pending approval and saved session ID.
ADK preserves the paused run. Agent Runtime stores the session after deployment, and Agent Identity gives the deployed agent its own cloud identity.
To get started I cloned the Agents CLI repository (Agents CLI GitHub Repo → https://fandf.co/3VZcs8S) and installed its seven skills.