Google released an open-source CLI and skills package called agents-cli that turns coding agents into deployment pipelines for Google Cloud. The tool, hosted on GitHub and documented at google.github.io/agents-cli, bundles seven skills that teach Antigravity CLI, Claude Code, Codex, or any other coding agent how to scaffold, evaluate, deploy, and monitor AI agents built with Google’s Agent Development Kit (ADK).
The move is Google’s bid to own the deployment layer of the agent-building stack. Every major AI lab ships coding agents — Anthropic’s Claude Code, OpenAI’s Codex, and Google’s own Antigravity CLI — but none of them natively knows how to push an agent to production on a specific cloud. agents-cli fills that gap with a single command: uvx google-agents-cli setup installs the CLI and context-aware skills. After that, a developer can ask Claude Code to “build a support agent that answers questions from our docs,” and the coding agent uses the installed skills to scaffold the project, write ADK Python code, generate evaluation datasets, grade the outputs, and deploy to Cloud Run or GKE.
The architecture is telling. agents-cli is not a coding agent itself. “No. agents-cli is a tool for coding agents, not a coding agent itself,” the FAQ states. It is a skill layer that sits between the coding agent and the cloud platform. The seven skills cover the full lifecycle: workflow (development lifecycle and code preservation rules), ADK code (the ADK Python API for agents, tools, orchestration, callbacks, and state), scaffold (project creation and enhancement), eval (metrics, datasets, LLM-as-judge grading, adaptive rubrics, failure-mode clustering), deploy (Agent Runtime, Cloud Run, GKE, CI/CD, secrets), publish (Gemini Enterprise registration), and observability (Cloud Trace, logging, third-party integrations).
The eval commands are the most interesting part of the release. agents-cli includes agents-cli eval generate to run an agent on an evaluation dataset and produce traces, agents-cli eval grade to score those traces against metrics, agents-cli eval compare to diff two eval result files, agents-cli eval analyze to cluster failure modes from grade results, agents-cli eval optimize to auto-tune agent prompts using eval data, and agents-cli eval dataset synthesize to generate multi-turn evaluation scenarios. That is a full evaluation pipeline packaged as CLI commands that a coding agent can invoke on its own output.
This is Google’s playbook from Kubernetes and TensorFlow: build the open-source tool that becomes the standard interface for a new category. With Kubernetes, Google gave the industry a standard way to orchestrate containers. With TensorFlow, it gave researchers a standard framework for deep learning. With agents-cli, Google is trying to give the industry a standard CLI for the agent lifecycle — one that happens to deploy to Google Cloud.
The bet is that developers will choose the path of least resistance. If a coding agent can scaffold an agent project, evaluate it, and deploy it with a few natural-language prompts, the developer has little incentive to learn the equivalent AWS or Azure tooling. Google’s cloud competitors will need to ship their own skill packages, or the agent-deployment standard will default to Google Cloud by convenience.
The timing matters. Agent-building is still in its early majority phase. Most developers who build agents today do so manually — they write Python code with LangChain or the ADK directly, run evaluations in notebooks, and deploy with custom CI/CD pipelines. agents-cli abstracts all of that into a skill that a coding agent can invoke. For a developer who already uses Claude Code or Codex, the friction to try Google Cloud drops to nearly zero.
Google also made the tool work without a coding agent. The CLI commands run standalone from the terminal: agents-cli scaffold my-agent creates a new project, agents-cli eval generate runs evaluations, agents-cli deploy pushes to production. The skills are an accelerator, not a requirement. That dual-mode design lowers the adoption bar for teams that are not yet using coding agents.
The open-source licensing and the GitHub-hosted repository signal Google’s intent to build community around the tool. The README invites bug reports and feature requests via issues, and the team asks builders to share projects at [email protected]. The contributing guide is short — feedback is the primary contribution model — but the repository structure leaves room for community extensions.
The biggest open question is whether agents-cli becomes a standard or a Google Cloud lock-in mechanism. The tool is open source, but it deploys exclusively to Google Cloud services: Agent Runtime, Cloud Run, GKE, and Gemini Enterprise. There is no AWS or Azure target in the current command set. Google is betting that the convenience of a single CLI that handles the full lifecycle will outweigh multi-cloud flexibility for most teams.
The seven skills cover the agent lifecycle end to end, from scaffold to observability. That completeness is the product.