Kipt

Your prompts and skills, one keystroke away.

Explore

google-agents-cli-workflow skill: the agent build lifecycle

The agents-cli workflow skill from google/agents-cli walks coding agents through building ADK agents: spec, scaffold, build, eval, deploy, publish and observe.

By the Kipt team · Updated 26 October 2026

Source: google/agents-cli · Maintainer: Google · License: Apache-2.0 · Checked: 26 October 2026

What is the google-agents-cli-workflow skill?

google-agents-cli-workflow is the entry point to Google's agents-cli skill set. It gives a coding agent the whole lifecycle for building an AI agent, from the first design question to production monitoring, and tells it which of the other six skills to load at each phase. Its SKILL.md description calls it the "Entrypoint for building ADK agents."

It lives in google/agents-cli, which the README describes as "The CLI and skills for building agents on Gemini platform." The README says the tool works with Antigravity CLI, Claude Code and Codex "and any other coding agent," and that agents-cli "is a tool for coding agents, not a coding agent itself." The default framework is ADK (Agent Development Kit).

As of 26 October 2026, its skills.sh page listed about 511,900 installs (snapshot, 26 October 2026). The SKILL.md metadata lists Google as author and version 1.11.0, and it asks for google-agents-cli ~= 1.11.0.

What does it do?

The SKILL.md is a phase-by-phase playbook. It tells the agent to re-read the matching skill before each phase, because "Context compaction may have dropped earlier skill content."

Eight phases, seven skills

Phase What happens Skill to load
0. Understand Design dialogue, write .agents-cli-spec.md none
1. Study recipes Check the recipe index for a matching sample google-agents-cli-adk-code
2. Scaffold agents-cli scaffold create or scaffold enhance google-agents-cli-scaffold
3. Build Write agent code, smoke test with agents-cli run google-agents-cli-adk-code
4. Evaluate agents-cli eval run, fix, repeat google-agents-cli-eval
5. Deploy agents-cli deploy after explicit approval google-agents-cli-deploy
6. Publish Optional registration with Gemini google-agents-cli-publish
7. Observe Traces, logging, monitoring google-agents-cli-observability

A spec before any code

Phase 0 follows references/brainstorming.md: ask one question at a time, propose 2 to 3 architecture approaches for non-trivial agents, and get approval. The topics are the problem, external APIs and auth, safety constraints and deployment preference. The SKILL.md says: "Do NOT proceed to planning, scaffolding, or coding until the user approves the spec." A trivial agent with a single tool and a fixed persona gets a couple of questions, a spec of 2 to 3 sentences and one approval.

Recipes over reinvention

Phase 1 sends the agent to a topic index of reference recipes for capabilities such as retrieval over your own documents, sandboxed code execution, memory across conversations, approval gates, OAuth consent and scheduled runs. The skill's view is blunt: skipping a matching recipe "means rebuilding patterns that already exist, usually worse."

Eval is the main gate

Phase 4 is called "the most important phase." The skill separates three kinds of checks: unit tests for code correctness, agents-cli eval for agent behavior, and agents-cli run "prompt" for quick smoke tests. It tells the agent to start with 1 to 2 eval cases and expect 5 to 10 or more iterations. It also notes that eval run exits 0 whatever the scores are, so the agent has to read them.

Operating rules

The second half covers behavior rather than phases:

  • Code preservation. Change only the targeted lines and keep the model, version, API key settings, comments and formatting as they were.
  • Model choice. "NEVER change the model unless explicitly asked." For new agents, it gives a short google-genai script that lists the models available in your location.
  • Loops. Stop if the same error appears 3 or more times, and use terraform import for 409 state conflicts instead of retrying.
  • Debugging. Reproduce, localize, fix one thing, verify, then guard with an eval case.
  • Shortcuts to resist. A table of excuses such as "I can skip the scaffold and set up manually," each with why it fails.

Six reference files support it: brainstorming.md, spec-template.md, terminology.md (product names to CLI values), commands.md, internals.md and extension.md, which explains how to override or add agents-cli commands.

How do you install it?

Google's Getting Started guide lists Python 3.11+, uv and Node.js as prerequisites, with the Google Cloud SDK and Terraform as optional extras for deployment. The recommended install adds the CLI and all seven skills:

uvx google-agents-cli setup

Or install skills only, either the whole set or just this one (the command shown on skills.sh):

npx skills add google/agents-cli
npx skills add https://github.com/google/agents-cli --skill google-agents-cli-workflow

Because the workflow skill hands off to the others at each phase, installing the full set makes more sense than this skill alone. The skill's metadata lists the agents-cli binary as a requirement (uv tool install google-agents-cli). The guide supports macOS, Linux and Windows through WSL 2.

In Claude Code, run /skills; the guide says you should see google-agents-cli-workflow listed. Codex and Cursor read the same SKILL.md format; see Codex skills and Cursor skills.

How do you use it?

Its description says it is "Always active" and should load when you ask to "develop an agent", "build an agent using ADK", "run the agent locally", "debug agent code" or similar. You can also call /google-agents-cli-workflow in Claude Code. The README's own first example is: "Use agents-cli to build a caveman-style agent that compresses verbose text into terse, technical grunts".

Example prompts (suggestions, not quotes from the repo):

  • Use agents-cli to build a support agent that looks up orders by ID. Ask me questions first.
  • I have an existing ADK agent in ./agent. Bring it into the agents-cli workflow without changing its logic.
  • Which reference recipes match an agent that needs memory across sessions and an approval step before refunds?
  • The agent works in agents-cli run but fails in the playground. Debug it step by step.
  • We are past eval. Walk me through what is left before production.

The Lifecycle page in Google's docs describes the same loop (scaffold, eval, deploy, observe) from the human side.

Tips and limits

From the SKILL.md, the repo docs and skills.sh, not from testing:

  • Expect questions first. The design dialogue is deliberate. If you want a fast prototype, say the agent is trivial or answer the questions briefly.
  • Google Cloud is optional at first. The README says local development (create, run, eval) works with an AI Studio API key; deployment needs Google Cloud.
  • Telemetry. The README says agents-cli sends one usage signal per command to your own Google Cloud project, with metadata in request headers that Google receives. Opt out with AGENTS_CLI_TELEMETRY=0 or DO_NOT_TRACK=1.
  • Stale skills. If skills seem outdated, the skill suggests agents-cli setup --skip-auth to reinstall them.
  • Security audits. On skills.sh, Gen Agent Trust Hub rated it Pass with risk level Safe (22 October 2026); its report notes the uv install, the extension mechanism and an indirect prompt injection surface from reading .agents-cli-spec.md, and credits the explicit approval gates. Socket (20 October 2026) and Snyk (risk level Low, 20 October 2026) rated it Pass.

Under Apache 2.0 you can copy and modify the skill as long as you keep the license and notices and mark files you change.

Save it in your Kipt library

In Kipt, open Skills and paste the skill's GitHub folder link to import it with its six reference files, or paste the repo link and pick from the list of all seven agents-cli skills. You can also drop the folder or a .zip. Your skills live in your own private GitHub repo and install into Claude Code or Codex with one command:

npx skills add https://github.com/<you>/<repo>.git --all -g

Kipt is not affiliated with Google.

Similar skills

Questions

What is the google-agents-cli-workflow skill?

It is the entry point of the seven Agent Skills in Google's google/agents-cli repo. It gives a coding agent the full lifecycle for building an agent with ADK or another framework: understand the goal, study recipes, scaffold, build, evaluate, deploy, publish and observe, plus rules for preserving code, choosing models and debugging.

Does the workflow skill make my coding agent ask questions before building?

Yes. Phase 0 tells the agent to run a design dialogue, asking one question at a time, and to write the agreed design to .agents-cli-spec.md. The SKILL.md says not to plan, scaffold or code until you approve the spec, and it scales the questions down for a trivial agent.

How do I install the agents-cli workflow skill?

Run uvx google-agents-cli setup to install the agents-cli command and all seven skills. To add only this skill, run npx skills add https://github.com/google/agents-cli --skill google-agents-cli-workflow. Prerequisites are Python 3.11 or later, uv and Node.js.

Why does the workflow skill tell my agent not to change the model?

Its first principle is code preservation: change only what you asked for and keep everything else, including the model setting. The skill says the scaffolded model was chosen deliberately and that changing it often breaks things, such as a wrong location, a deprecated version or a 404.

Should my agent write unit tests for the LLM's answers?

No. The skill says never to write unit tests that check LLM response content, because the output is non-deterministic. Use agents-cli run for quick smoke tests and agents-cli eval with LLM-as-judge criteria for behavior, and keep pytest or go test for code correctness.

More in Skill directory

All guides

Keep your prompts and skills in one place, in your own GitHub. Install every skill with one command.

Get started

Kipt is an independent product, not affiliated with or endorsed by Anthropic, OpenAI, Cursor or GitHub. Claude, ChatGPT, Codex, Cursor and GitHub are trademarks of their owners.