On August 3, 2026, Google shipped Agents CLI in Agent Platform. Skim the launch and it reads like one more terminal coding agent in a crowded field. Read the actual tagline and it's something stranger and more strategic: "the CLI and skills that turn any coding assistant into an expert at creating, evaluating, and deploying AI agents on Google Cloud." It is not competing with Claude Code. It's climbing inside it.

The one-screen answer#

It's skills, not an agent — and that's the whole point#

The install command is the tell. You don't launch agents-cli and start chatting. You run one setup command and it drops skill files into whatever coding agent you use. From then on, when you tell Claude Code "scaffold an ADK agent that answers billing questions, evaluate it, and ship it," your agent already knows the ADK API patterns, the eval methodology, and the exact deploy invocation — because Google wrote those seven skills and your agent now reads them.

This is the same SKILL.md pattern we've covered before: a skill is a scoped instruction file a coding agent loads on demand, and whether it fires at all comes down to the description. Google isn't shipping a product you learn. It's shipping competence your agent absorbs — workflow, adk-code, scaffold, eval, deploy, publish, observability — so the human never has to open the Cloud console.

Google isn't shipping a product you learn. It's shipping competence your existing agent absorbs — so the deploy target becomes a reflex, not a decision.

The wedge#

Step back and the strategy is clean. In the agent stack, three layers are contested: the model, the framework, and the deploy target. Google can't force its model on you (you'll use whatever's cheapest per token this week) and can't force its framework (ADK competes with LangGraph, Pydantic AI, and the OpenAI Agents SDK). But the deploy target is decided by muscle memory — by whatever your tooling reaches for when you say "put this in production."

Agents CLI plants that muscle memory inside the coding agent itself. When "now deploy it" resolves, without a decision, to agents-cli deploy targeting Agent Runtime, GCP has won the layer that actually bills. AWS and Azure own their consoles; a founder has to go there. Google's move skips the console entirely and rides in on the CLI you already had open. It's the cheapest adoption path anyone has shipped for an enterprise agent runtime — and, not coincidentally, the stickiest.

What a founder should actually do#

Two honest caveats keep this from being a pure win:

  1. It's Google Cloud only. The skills teach exactly one deploy story. If you're multi-cloud or cloud-agnostic by principle, installing them nudges every future "deploy" toward GCP. That's the lock-in the convenience is paying for.
  2. You don't need the skills to get the CLI. agents-cli scaffold, eval generate, deploy, and publish gemini-enterprise all run standalone from your terminal. If you want Google's deploy path without wiring it into your agent's reflexes, script the commands directly and skip the skills injection.

If your team already lives in a coding agent all day and you've made peace with GCP, this is the lowest-friction path to a governed, observable agent runtime that exists right now — install the skills and let the agent drive. If you're deliberately keeping your deploy target portable, treat Agents CLI as a standalone CLI you call explicitly, not a set of reflexes you hand your agent. Either way, the launch is worth reading as the opening move it is: the cloud war for agents is being fought one skill file at a time.

For the hands-on version — empty folder to a deployed agent, command by command — see our companion walkthrough: From Empty Folder to Deployed Agent: Google's Agents CLI, Step by Step. And if you're choosing which coding agent to install these into, start with Claude Code vs Codex CLI vs Gemini CLI.