Agents CLI – Google Cloud's Official Command-Line Tool for AI Agents

1. What Is Agents CLI

Agents CLI is Google Cloud's official command-line tool for simplifying the full AI Agent lifecycle. It packages more than a decade of Google Cloud engineering into seven structured skill modules covering scaffolding, ADK development, local evaluation, deployment to Cloud Run or GKE, and Gemini Enterprise publishing. Developers can let AI coding assistants run the full flow automatically or take over with manual CLI commands. Local development needs only an AI Studio API key—no GCP project required for prototyping—lowering the bar for production-grade agent work.

Image source: Official article

Technical positioning and domain: Agent engineering × DevOps—standardized flow from development to deployment by merging cloud infrastructure with the Agent Development Kit (ADK), not just model invocation or a bare framework.

Development background: Built by the Google Cloud Agent Platform team on Gemini, ADK, and A2A (Agent-to-Agent) expertise. Motivation: customers struggle turning prototypes into production—manual config, scattered docs, no standard toolchain.

Core value: Resolves "easy prototype, hard production." Prototypes take hours; production deploy, monitoring, eval pipelines, and CI/CD take days or weeks. Skill-pack injection embeds engineering knowledge in coding assistants so teams ship without deep cloud expertise—compressing days to hours.

Technical characteristics: ADK + A2A orchestration for interoperable agents; layered cloud-native design (orchestration, models, data, observability). Innovation: context compression—connection and deploy rules pre-encoded as structured commands so assistants don't re-infer infra each turn—lower tokens, higher determinism.

2. Key Features

• Skill-pack injection: Seven modules—Workflow, ADK Code, Scaffold, Eval, Deploy, Publish, Observability—installed via npx skills add into Gemini CLI, Claude Code, Codex, etc., so assistants generate accurate agent code and configs without guessing cloud wiring.

• Project scaffolding: scaffold creates standard agent projects (layout, ADK starter, deps, defaults) or adds deploy/CI/CD/RAG to existing repos and upgrades to new conventions.

• Local dev and debug: run for smoke tests; playground Web UI at http://127.0.0.1:8000; install and lint for deps and quality—local loop complete.

• Evaluation pipelines: eval run on datasets with metrics; eval compare across versions; LLM-as-Judge and trajectory scoring for output quality and behavioral consistency pre-deploy.

• Multi-target deploy: deploy to Agent Runtime (memory + orchestration), Cloud Run (serverless), or GKE (scale)—sensible defaults without deep container expertise.

• Enterprise publish: publish gemini-enterprise registers agents for org-wide search and reuse—key for large enterprises.

• Infrastructure as code: infra generates Terraform, staging/prod CI/CD, vector stores, and RAG data pipelines—infra in the standard agent workflow.

3. How to Use

1. Environment: Python 3.11+, uv, Node.js (skills). Linux/macOS recommended; Windows via WSL2. Verify: python --version && uv --version && node --version.

2. Install CLI + skills: uvx google-agents-cli setup—installs CLI and deploys all seven skill packs to your assistant.

3. Skills only (optional): npx skills add google/agents-cli if CLI is managed elsewhere.

4. Auth: agents-cli login --interactive for GCP browser login. Local: export GEMINI_API_KEY="your-key"—no GCP project for prototyping.

5. Create project: agents-cli scaffold my-first-agent—standard agent/, tests/, config/, etc.

6. Install deps: agents-cli install from requirements.txt or pyproject.toml.

7. Local debug: agents-cli playground at http://127.0.0.1:8000—pick agent, chat, validate logic.

Best practices: Run agents-cli info first; commit scaffold output for team consistency.

4. Pros and Cons

5. Comparison with Similar Tools

Selection advice: Google Cloud shops → Agents CLI first—deep integration with Cloud Run, GKE, Gemini Enterprise. Local-only coding → Cursor CLI or LangChain CLI. Non-GCP clouds → other platforms; deploy stays GCP-bound.

6. Editor's Take

The standout idea is skill-pack injection—not just commands, but encoded engineering knowledge in assistant context. "Knowledge as code" teaches when and why to run deploy/eval/infra steps—a pattern worth watching beyond Google.

Agents CLI hits the "last mile" after LangChain-style prototypes: standardized eval, deploy, observability. Trajectory scoring gives quantifiable pre-production quality—critical in prod.

Audience: GCP enterprise DevOps/AI engineers; individuals can prototype locally but deploy stays on Google Cloud.

Future: broader A2A tooling, maybe optional non-GCP targets; open custom skill packs could grow the ecosystem.

— innovation 4.8, utility 4.5, ease 4.3, community 3.8; GCP dependency is the main caveat.

7. Use Cases

• Enterprise multi-agent systems: Scaffold + eval + deploy for customer service, workflow engines.

• Rapid prototyping: AI Studio key + playground for hours-to-demo validation.

• Assistant augmentation: Skill injection so Cursor/Claude Code write Terraform and CI/CD correctly.

• DevOps IaC: infra for staging/prod Terraform, vector stores, RAG pipelines.

• Org agent catalog: publish gemini-enterprise for internal discovery and reuse—AI CoE teams.

8. FAQ

Q: Deploy outside Google Cloud?
A: Not currently—only Agent Runtime, Cloud Run, GKE. AWS/Azure → LangChain CLI or native cloud tools.

Q: Local dev without GCP project?
A: Yes—GEMINI_API_KEY from AI Studio suffices until production deploy.

Q: Do skill packs break other assistant features?
A: No—they add context; default behavior remains for non-agent tasks.

Q: vs. LangChain?
A: Complementary—Agents CLI uses ADK, not LangChain core, but can wrap LangChain tools via ADK Python APIs.

Q: Custom eval metrics?
A: Extend eval config or custom callbacks; docs thin—see GitHub examples.

Q: Windows?
A: Supported; WSL2 recommended for paths and Node skill install.

Q: Custom skill packs?
A: Official packs only for now; future openness possible—workaround via assistant config/rules.

9. Project Links

• Docs: https://google.github.io/agents-cli/
• GitHub: https://github.com/google/agents-cli
• Google Cloud: https://cloud.google.com/agent-platform/docs/agents-cli-overview