Google Adds Framework Extensions to Agents CLI

Google Cloud announced that Agents CLI version 1.5.0 supports extensions, allowing developers to replace built-in commands or add commands without forking the tool. Version 1.8.0 includes a first-party LangChain extension that substitutes LangChain commands for framework-specific ones while leaving the rest of the agent development lifecycle intact. The extension system is experimental in v1.8.0, and its manifest and commands may change in breaking ways; teams using it should pin their Agents CLI version.

The extension system is intended for teams using frameworks such as LangChain, LangGraph, or Deep Agents, as well as enterprise platform teams enforcing shared policies. Extensions are directories in Git repositories containing an agents-cli-extension.yaml file, which lists command overrides and added commands. They can be referenced from Google’s repository, another Git repository, or a local path; a repository can hold multiple extensions selected with a #name fragment. Installation can be scoped to a project, where the extension and lockfile entry are committed with the project, or to a user with --global, which applies across projects on one machine. If project and user scopes define the same command, the project scope takes precedence.

The lockfile records the exact Git commit SHA resolved during installation, so agents-cli install restores the pinned code on fresh checkouts without advancing the pin. Developers use agents-cli extension update to resolve a newer commit, which prompts them again to trust third-party code. Google says the system addresses a limitation in framework-specific toolchains: teams otherwise may have to maintain a CLI fork, rely on wrapper scripts that can be bypassed, or build replacement tooling from scratch, potentially spending several engineer-months. In Agents CLI, six commands assume the ADK runtime—playground, run, and four eval subcommands—while about eighteen others, including scaffold, infra, deploy, publish, and login, are framework-agnostic; without extensions, a non-ADK project could lose access to all twenty-four.

Google’s LangChain template scaffolds an agent project that uses the Agents CLI lifecycle for local runs, deployment, and evaluation, with no google-adk dependency. It also supports LangGraph graphs and Deep Agents: the article says LangChain 1.0 and later uses LangGraph for state management, and all three frameworks share a graph representation. The walkthrough requires Python 3.11 or later, the uv package manager, Google Cloud SDK, and a Google Cloud project with billing enabled. Developers create the project using the template pinned to v1.8.0, then run agents-cli install. The generated project includes an extension manifest, a LangGraph development-server configuration, application code, evaluation files, Terraform deployment modules, and a uv.lock dependency lockfile.

The article also demonstrates a custom extension that wraps the deploy command with a Software Bill of Materials (SBOM) scan, a list of software components and dependencies. It presents this as a pattern for platform teams to apply policy checks across projects, alongside examples such as compliance scans, corporate single sign-on for login, and Model Risk Management steps for evaluations. For an agent example, the template creates a Gemini ReAct agent with a simulated weather tool and sets the model to gemini-3.8-flash. Google says its extension mechanism lets teams adopt the deployment, evaluation, and observability capabilities of Agents CLI while choosing a different framework or changing command behavior.

Google Cloud announced that Agents CLI version 1.5.0 supports extensions, allowing developers to replace built-in commands or add commands without forking the tool. Version 1.8.0 includes a first-party LangChain extension that substitutes LangChain commands for framework-specific ones while leaving the rest of the agent development lifecycle intact. The extension system is experimental in v1.8.0, and its manifest and commands may change in breaking ways; teams using it should pin their Agents CLI version.

The extension system is intended for teams using frameworks such as LangChain, LangGraph, or Deep Agents, as well as enterprise platform teams enforcing shared policies. Extensions are directories in Git repositories containing an agents-cli-extension.yaml file, which lists command overrides and added commands. They can be referenced from Google’s repository, another Git repository, or a local path; a repository can hold multiple extensions selected with a #name fragment. Installation can be scoped to a project, where the extension and lockfile entry are committed with the project, or to a user with --global, which applies across projects on one machine. If project and user scopes define the same command, the project scope takes precedence.

The lockfile records the exact Git commit SHA resolved during installation, so agents-cli install restores the pinned code on fresh checkouts without advancing the pin. Developers use agents-cli extension update to resolve a newer commit, which prompts them again to trust third-party code. Google says the system addresses a limitation in framework-specific toolchains: teams otherwise may have to maintain a CLI fork, rely on wrapper scripts that can be bypassed, or build replacement tooling from scratch, potentially spending several engineer-months. In Agents CLI, six commands assume the ADK runtime—playground, run, and four eval subcommands—while about eighteen others, including scaffold, infra, deploy, publish, and login, are framework-agnostic; without extensions, a non-ADK project could lose access to all twenty-four.

Google’s LangChain template scaffolds an agent project that uses the Agents CLI lifecycle for local runs, deployment, and evaluation, with no google-adk dependency. It also supports LangGraph graphs and Deep Agents: the article says LangChain 1.0 and later uses LangGraph for state management, and all three frameworks share a graph representation. The walkthrough requires Python 3.11 or later, the uv package manager, Google Cloud SDK, and a Google Cloud project with billing enabled. Developers create the project using the template pinned to v1.8.0, then run agents-cli install. The generated project includes an extension manifest, a LangGraph development-server configuration, application code, evaluation files, Terraform deployment modules, and a uv.lock dependency lockfile.

The article also demonstrates a custom extension that wraps the deploy command with a Software Bill of Materials (SBOM) scan, a list of software components and dependencies. It presents this as a pattern for platform teams to apply policy checks across projects, alongside examples such as compliance scans, corporate single sign-on for login, and Model Risk Management steps for evaluations. For an agent example, the template creates a Gemini ReAct agent with a simulated weather tool and sets the model to gemini-3.8-flash. Google says its extension mechanism lets teams adopt the deployment, evaluation, and observability capabilities of Agents CLI while choosing a different framework or changing command behavior.