Agents CLI: Google turns your editor into an ADK specialist

TL;DR Agents CLI (announced by Addy Osmani at Google Cloud Next) installs 7 skills in your coding assistant — Claude Code, Codex, Antigravity — covering the full lifecycle of an ADK agent: scaffolding, workflow design, Agent Runtime deployment, Sandbox, MCP/A2A wiring, evaluation and surfacing in Gemini Enterprise. The shortest path today from "I want to try agents on the Cloud" to "I deployed a real agent".

The hard part about building an agent in 2026 is not writing SDK code — it is knowing which of the 40+ Agent Platform components to use, in what order, with what configuration. A generic coding assistant guesses. Agents CLI teaches it to make the calls a seasoned Agent Platform engineer would make.

The announcement is from Addy Osmani (Director, Google Cloud AI) at Google Cloud Next — complementing the Agent Platform launch on April 22.

The 7 skills Agents CLI installs

A single install injects a cohesive skill pack into your editor. They cover the agent from scaffold to surfacing — all aligned with the official ADK + Gemini Enterprise architecture.

1 · Bootstrap

🏗️ Project scaffolding

Correct initial layout — folder structure, pyproject.toml, ADK config, compatible version pins.

2 · Architecture

🧩 ADK workflow design

Pick between sequential, collaborative (multi-agent) or dynamic graph — with rationale for when each fits.

3 · Execution

🚀 Agent Runtime deployment

Valid Runtime config with sub-second cold start, scaling and observability on day 1.

4 · Security

🛡️ Agent Sandbox integration

Least-privilege permissions by default; scoped tokens, isolated execution of generated code.

5 · Connectivity

🔌 Tool wiring

Correct integration of MCP, A2A and native connectors — no more pasting tutorial YAML.

6 · Quality

📊 Offline and online evaluation

Pre-seeded evaluation dataset + production metrics instrumentation to avoid silent regressions.

7 · Delivery

✨ Surfacing in Gemini Enterprise

Publishes the agent inside the Gemini Enterprise app — where the end user already is, without a custom front-end.

Before vs. after Agents CLI

MCP docs, ADK docs, gcloud docs, Runtime docs — today they live in four places. Agents CLI collapses that at the moment it matters: when the developer is asking the editor to generate code.

Decision	Before (generic editor)	After (Agents CLI)
Which ADK graph type?	You pick by trial; the editor accepts anything	Skill recommends sequential / collaborative / dynamic with rationale
How to scope Sandbox permissions?	Copy tutorial, usually too broad	Least-privilege by default, tokens scoped to the use case
How to evaluate before production?	Skip the step or improvise manual tests	Pre-seeded dataset + offline and online evaluation suite
Where to publish to end users?	Custom Streamlit or equivalent front-end	Native surfacing in the Gemini Enterprise app
MCP, A2A and connectors	Four doc pages open across four tabs	One skill guides when to use each protocol

The concrete flow: "scaffold a customer-support agent"

Addy describes the use loop with a real prompt. Editor with Claude Code or Antigravity attached, you type:

"Scaffold a customer-support agent that calls our order-status MCP server and logs to Observability."

The skills fire in sequence — you review and commit:

Correct project layout

Standard ADK structure, pinned dependencies, documented environment variables.

Well-formed ADK graph

Explicit choice between sequential, collaborative or dynamic — with a rationale paragraph in the README.

Valid deployment config

Agent Runtime wired for cold start, scaling, Observability plugged in from the start.

Least-privilege Sandbox

The agent only sees the order-status MCP; no extra permissions granted "just in case".

Pre-seeded evaluation dataset

Canonical customer-support cases already in the repo — you leave zero with a measurable baseline.

What the design gets right

Three architectural choices that separate "yet another scaffolder" from a tool teams will actually adopt:

Coding-assistant agnostic. Works with Claude Code, Codex, Antigravity — the editor choice stays yours. No extra lock-in.
Discovery-first. Skills explain why a decision was made, not just what. The assistant teaches while it builds — your team gets better at Agent Platform on the job.
Reduces fragmented-docs friction. Knowledge arrives at the right moment, inside the editor, not across four parallel tabs.

The goal is that your coding assistant teaches as it builds — so the humans on your team also get better at Agent Platform.

Honest caveats before you adopt

⚠️ What the comments on the announcement surfaced Public discussion of Agents CLI raised four legitimate concerns no scaffolder resolves on its own: token consumption (richer skills grow the context on every call), drift across variants (different teams ship agents with diverging conventions), edge-case debugging (when the decision is the skill's, the root cause is less obvious) and architectural accountability (who owns the graph choice when the agent picked it?).

None of this invalidates the tool — but it sets the right contract: Agents CLI is a decision accelerator, not a replacement for architectural review. In Autenticare projects we keep a rule: the skill suggests, the human approves, Git records.

Where to start

Install Agents CLI in your favorite editor (Claude Code, Codex or Antigravity).
Scaffold a small agent — a low-risk internal case, ideally read-only on the first iteration.
Read the rationale the skill wrote in the README before committing. That is the part that teaches.
Enable Observability before any write to a production system.
Add your own evaluation dataset alongside the pre-seeded one — to measure your case, not the generic one.

Autenticare Diagnosis

Want your first ADK agent in production in 30 days?

We pick the use case, apply Agents CLI with your team's editor, define governance (Sandbox, evaluation, Observability) and go live on Gemini Enterprise. Zero boilerplate, maximum knowledge transfer.

Book diagnosis Calculate ROI