Powering the Agent Development Lifecycle with any framework, from ADK to LangChain and beyond

Summary

Google's Agents CLI scaffolds, runs, evaluates, and deploys agents on Google Cloud. Its extension system (added in 1.5.0, labelled experimental in 1.8.0) lets teams override built-in commands or add new ones. The first official extension is a LangChain template, which lets a LangChain or LangGraph agent use the same scaffold, run, and deploy workflow without the Agent Development Kit (ADK).

Key points

• Install: uvx google-agents-cli setup.
• LangChain agent on Cloud Run: agents-cli create my-agent --agent google/agents-cli/extensions/langchain/template@v1.8.0 -d cloud_run, then agents-cli install, agents-cli run "...", agents-cli playground (a LangGraph dev server on port 8080), and agents-cli deploy. The ADK-only commands eval dataset and eval optimize refuse to run and explain why.
• Extensions: an extension is a directory in any Git repository with an agents-cli-extension.yaml manifest. There is no central registry. Add one with agents-cli extension add <owner/repo#name> (or local@<path>). Scope is per project (default, committed) or --global, and project scope wins.
• Pinning: a lockfile pins each extension to a commit SHA, and agents-cli extension update moves the pin forward. Help output marks overridden commands with [↑ <extension>].
• Platform policy: the example sbom-deploy extension runs syft and grype before the native deploy and fails it on critical findings. Extensions call the native command with AGENTS_CLI_DISABLE_OVERRIDES=1 to avoid loops.

Why it matters

Google is decoupling its agent deployment tooling from ADK. Teams can keep LangGraph and still use Google's scaffold, evaluation, and deploy workflow. Platform teams also get a supported way to enforce checks such as SBOM scans in every deploy.