Google's agents-cli adds agent-building and eval skills to Claude Code and Codex

Google's agents-cli installs seven skills that teach coding agents to scaffold, evaluate and deploy ADK agents, with deployment aimed at Google Cloud.

Google's agents-cli installs seven skills into the coding agent you already use, including Antigravity CLI, Claude Code and Codex. The skills cover ADK code, scaffolding, evaluation, deployment, publishing and observability. For builders, the eval skill is the useful part. The repo says it teaches metrics, datasets, LLM-as-judge and adaptive rubrics, and it comes with commands such as agents-cli eval run, eval analyze, which clusters failure modes, and eval optimize, which auto-tunes prompts from eval data.

Setup needs Python 3.11+, uv and Node.js. You run uvx google-agents-cli setup, or npx skills add google/agents-cli for the skills only. According to the repo FAQ, local create, run and eval work with an AI Studio API key. Deployment and cloud features require Google Cloud. The CLI also sends usage telemetry by default, and you can turn it off with AGENTS_CLI_TELEMETRY=0.

The skills are built around ADK. Google's developer blog says an extension system arrived in version 1.5.0 and a first-party LangChain extension in 1.8.0.

@_avichawla writes that he used it in Claude Code to build a RAG agent that pauses for human approval, exits, and resumes from a new process. He reports that retrieval ran once and the answer was published once. That pause-and-resume behavior comes from ADK's tool confirmation and ResumabilityConfig, which he configured himself.

That path still has rough edges. Cadenya notes that ADK's docs list DatabaseSessionService and VertexAiSessionService as not supported for tool confirmation. Open ADK issues also report that paused confirmations get dropped inside AgentTool and that a paused sub-agent hijacks later turns.

Sources