Google's Agents CLI: A Terminal Path to Agent Platform
Google Cloud has introduced Agents CLI, a command-line tool that gives AI coding assistants a machine-readable interface for agent scaffolding and deployment.
On April 22, 2026, Google Cloud released Agents CLI, a specialized command-line interface integrated into the Gemini Enterprise Agent Platform. The tool is engineered to provide AI coding assistants, including Gemini CLI, Claude Code, and Cursor, with a machine-readable interface to the complete Google Cloud agent stack. For developers managing the Agent Development Lifecycle (ADLC), this allows local development environments to drive production deployments autonomously.
Installation and Bundled Skills
The interface is initialized through a core setup command, invoked via uvx google-agents-cli setup. To mitigate token waste and context overload caused by assistants reading extensive documentation, Google bundles direct API references as Skills. Developers inject these into their local coding environment using npx skills add google/agents-cli.
This package provides the AI assistant with direct sensory input and reference material for the Agent Development Kit (ADK). Utilizing agent skills correctly ensures the AI assistant maintains focus on implementation rather than searching for syntax rules.
Execution Modes and Deployment Targets
Agents CLI operates in two distinct modes. Agent Mode is optimized strictly for AI assistants to automate processes like unit testing and evaluation. Human Mode provides deterministic execution for human developers requiring manual terminal control.
The CLI automates Infrastructure as Code (IaC) injection and continuous deployment across three primary production targets:
| Deployment Target | Description | Infrastructure Requirement |
|---|---|---|
| Agent Runtime | Deploys source code as a tarball to Vertex AI reasoning engines. | Bypasses Dockerfile requirements. |
| Cloud Run | Deploys agents to standard containerized infrastructure. | Requires container configuration. |
| GKE Autopilot | Deploys agents to managed Kubernetes environments. | Built for enterprise-scale workloads. |
Ecosystem Integration
The release aligns with the broader consolidation of Vertex AI into the Gemini Enterprise Agent Platform. The CLI registers agents natively using agents-cli publish. It orchestrates Subagents, allowing developers to modularize specific tasks like codebase investigation within the framework. Structuring and configuring subagents in Gemini CLI is required to handle isolated, complex operations before compilation.
Agents CLI and the ADK ship with native support for the Model Context Protocol, allowing agents to discover system resources dynamically through Agent Cards. While optimized for Google Cloud infrastructure, the framework maintains multi-cloud flexibility. It includes LiteLLM as an escape hatch, permitting agents built via the CLI to execute inference using models like GPT-4o and Claude 3.5 Sonnet.
Governance and Security Controls
Runtime security is handled natively by Agent Gateway and Model Armor. These components enforce prompt injection defenses and data leakage policies automatically upon deployment. Technical analysis of the framework indicates that while the native governance is robust, Agent Gateway currently lacks a customer-pluggable webhook layer for external third-party policy enforcement.
If you build agents using AI coding assistants, transition your setup routines to utilize Agents CLI and its bundled skills. This integration standardizes your infrastructure pipeline and reduces the context window overhead required to map local code to Google Cloud deployment targets.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How to Provision Google Colab GPUs From the Command Line
Learn how to install the Google Colab CLI, provision high-performance remote GPUs from your local terminal, and execute headless machine learning workflows.
Hugging Face Defines the Scaffold vs Harness Agent Architecture
Hugging Face has published a new technical glossary formalizing the structural differences between an AI agent's scaffolding and its execution harness.
Gemini 3.1 Flash-Lite Ships 1M Context at $0.25 Per Million
Google's lowest-latency Gemini model is now generally available, introducing variable thinking levels and a 1M token context window for high-volume routing.
Runpod Flash Removes Container Overhead for AI Inference
The open-source Flash Python SDK allows developers to convert local functions into auto-scaling serverless AI inference endpoints without Dockerfiles.
$40 Billion Anthropic Deal Trades Equity for 1M Google TPUs
Anthropic will receive $10 billion in upfront cash and up to 1 million Ironwood TPUs in a $40 billion infrastructure agreement with Google.