Google's Agents CLI brings end-to-end tooling for agentic engineering as described by Andrej Karpathy — covering scaffolding, evaluation, and deployment in a single workflow. The tool injects 7 ADK-specific skills into coding agents like Claude Code, Cursor, and Codex, enabling them to drive the full agent lifecycle from natural language prompts. A step-by-step walkthrough demonstrates building a RAG agent from scratch: scaffolding with ADK templates, local testing, running a 20-scenario eval suite with LLM-as-judge scoring, deploying to Google Cloud Agent Runtime, and registering to Gemini Enterprise for org-wide discoverability. The eval step — often skipped in agent tutorials — caught a hallucination edge case traced to a single instruction line.
