// the find
google/agents-cli
The CLI and skills that turn any coding assistant into an expert at creating, evaluating, and deploying AI agents on Google Cloud.
A CLI plus a bundle of skill files that teach coding agents (Claude Code, Codex, Antigravity) how to scaffold, evaluate, and deploy agents built with Google's Agent Development Kit (ADK) onto Google Cloud. It's aimed at teams already committed to ADK/Gemini Enterprise who want their coding assistant to handle the Cloud Run/GKE/Agent Runtime deployment plumbing instead of hand-rolling Terraform and gcloud commands.
The eval subsystem is more than a smoke test — generate/grade/compare/analyze/optimize subcommands do LLM-as-judge grading, cluster failure modes across traces, and auto-tune prompts from eval results. Scaffolding covers four languages (Python, Go, Java, TS) plus a LangChain extension, each with its own eval dataset, load test, and CI/CD template rather than one generic stub reused everywhere. Deployment isn't locked to a single managed surface: Agent Runtime, Cloud Run, and GKE are all first-class targets with infra-as-code generation (`infra single-project`, `infra cicd`) rather than just a `gcloud deploy` wrapper. The telemetry section is unusually specific about what's collected (command name, exit code, duration) versus what isn't (file contents, arguments), and gives a real opt-out env var instead of a vague privacy promise.
Real usage is tied to Google Cloud — local dev works on an AI Studio key, but deploy/publish/infra commands assume a GCP project, billing, and IAM, so this is packaging around GCP services more than a portable agent framework. The core interaction model has a coding agent reading skill markdown and then running `deploy`, `infra single-project`, or `publish gemini-enterprise` on your behalf — those commands touch cloud infra and billing, and there's no mention of a dry-run or confirmation gate before a skill-driven agent executes them. Surface area is large (7 skills, dozens of subcommands spanning eval/deploy/infra/publish/observability) for a project that's still young, with no stated compatibility guarantee between skill versions and the underlying ADK/Cloud APIs they generate code against. The Go/Java/TS templates still carry raw `{{cookiecutter.*}}` placeholder filenames in the repo tree, suggesting those scaffolds get less exercise than the Python path.