google-agents-cli-workflow

Start by clarifying requirements and writing a spec, then scaffold the project with agents-cli scaffold create. Implement the agent code, smoke test with agents-cli run, validate behavior with agents-cli eval, and deploy only after explicit approval.

Install it with uv tool install google-agents-cli, then run uvx google-agents-cli setup to install the companion skills. Check your installation and project configuration anytime with agents-cli info.

pytest tests code correctness such as imports and return types, while agents-cli eval tests agent behavior including response quality, tool usage, and safety compliance. Never write pytest assertions on LLM output content because responses are non-deterministic.

agents-cli supports Agent Runtime (formerly Vertex AI Agent Engine), Cloud Run, and GKE as deployment targets. Use agents-cli scaffold enhance with the --deployment-target flag to add deployment support to a prototype project.

Deployment failures often stem from missing permissions, wrong project configuration, or stale credentials. Run agents-cli login --status to verify auth, check agents-cli info for project config, and consult the deploy skill's troubleshooting section for target-specific fixes.

You should never change the model unless explicitly requested, since the scaffolded model was chosen deliberately and arbitrary changes can break location or version compatibility. For new agents, list available models with the google-genai client to pick the latest one.