Google Managed Agents: Skip the Infrastructure. Ship the Agent.

Before we talk about what Google Managed Agents is, let us talk about what it feels like to need one.

You have a task. It is not just a question. It is a job. Research ten competitors. Download their pricing pages. Extract the numbers. Compare them. Write a report. Save it somewhere.

You could do this yourself. It would take a few hours. Or you could hire someone to do it. But what if you could describe exactly what needs to happen, hand it to a system, and come back to the finished report?

That is the promise of an AI agent. Google just made building one dramatically simpler.

What Is an AI Agent?

An AI agent is not a chatbot. A chatbot answers questions. It reads what you type, generates a response, and stops.

An agent acts.

An agent can search the web. Run code. Read files. Write files. Execute a sequence of steps. Make decisions along the way. If a step fails, it tries another approach.

Think of a chatbot as a knowledgeable colleague sitting at a desk, answering whatever you ask.

An agent is more like a junior employee with a laptop, an internet connection, and a task list. You explain the goal. They figure out how to get there.

The Problem Before Managed Agents

Building an AI agent used to mean building a lot more than just the agent.

You needed a secure environment for it to run in. A place where it could execute code without breaking anything else. You needed to manage that environment: start it, stop it, clean it up. You needed to wire up the tools the agent would use: web search, code execution, file access. You needed to handle state, making sure the agent remembered what it did in step three when it got to step seven.

All of that is real engineering work, and none of it is the actual thing you wanted to build.

Google Managed Agents removes every layer except the last one: your agent logic.

What Google Announced at I/O 2026

At Google I/O 2026 on May 19, Google introduced the Managed Agents API as part of the Gemini Enterprise Agent Platform.

The core idea is this: you describe what your agent should do, what tools it can use, and what constraints it should follow. You make one API call. Google Cloud spins up a fully isolated, secure Linux environment for your agent, gives it the tools you specified, and runs it.

You do not manage the environment. You do not handle the infrastructure. You do not wire up the tools from scratch. Google handles all of that.

The underlying engine that powers Managed Agents is called Antigravity, specifically the model antigravity-preview-05-2026, which is Google's agent-optimised variant of Gemini, purpose-built for multi-step agentic workflows.

How It Works

The Sandbox

Every Managed Agent runs inside a sandbox, a private, isolated container in Google Cloud. Think of it as a dedicated mini-computer that exists solely for your agent to work in.

What the sandbox gives the agent:

A full Linux environment with a persistent file system
A bash terminal to run commands and scripts
Pre-installed tools: Python 3.11, Node.js 20, git, curl, pandas, numpy, and more
Network isolation by default, no external access unless you explicitly allow it
A lifespan of 7 days, automatically renewed with each new interaction

Each agent gets its own sandbox. Nothing leaks between agents or between users.

The Two APIs

Managed Agents API is split into two complementary parts:

Agents API (the setup): you use this once to create and configure your agent. Define its instructions, its tools, its skills, its constraints. This creates the agent and stores the configuration.

Interactions API (the runtime): every time you want the agent to do something, you call this. Send a task. The agent reasons, uses its tools, and returns the result. Each interaction can pick up where the last one left off, with the same sandbox and all the files from previous steps still intact.

How You Define Your Agent

Instead of writing orchestration code, you define your agent using two simple markdown files:

AGENTS.md: the agent's identity. What is it for? What can it do? What rules does it follow? What skills does it have access to? Think of this as the briefing document your agent reads before starting any task.

SKILL.md: a packaged capability you want the agent to be able to use. A skill is a reusable module, for example a skill that knows how to query a specific database, or generate a specific type of report.

These files are versionable. You can track them in git, update them, roll them back, just like any other code.

Where Managed Agents Fits in Google's Agent Stack

Google announced four levels of building with agents at I/O 2026. Managed Agents sits at level two.

The four levels share the same underlying infrastructure and communicate using the Agent-to-Agent (A2A) protocol, so an agent built with ADK can call a Managed Agent as a sub-agent, and vice versa.

What a Managed Agent Can Actually Do

| Capability | What the agent can do | |---|---| | Code execution | Run Python, Bash, Node.js scripts inside the sandbox | | Web search | Query real Google Search and fetch results | | URL context | Open URLs and extract content from web pages | | File system | Read, write, edit, and list files inside the sandbox | | MCP servers | Connect to custom tools via the Model Context Protocol | | Persistent state | Files and installed packages survive across multiple interactions |

What It Cannot Do Right Now

As of the date this blog was written (June 2026), Managed Agents is in Preview, which means pre-general-availability. These limitations reflect what the product is today. By the time you are reading this, some or all of them may no longer apply.

It is not intended for production use or commercial applications
It should not be used with sensitive, proprietary, or confidential data
The sandbox does not provide real Google Cloud credentials, so the agent cannot directly access GCP services like Cloud Storage on its own
It is subject to change without notice and is not covered by any SLA
Model usage is billed at standard Vertex AI rates, there is no free usage tier

This is a product in active development. Google has been clear that it is available for limited testing and evaluation, not for shipping products to end users. That will change as it matures.

A Simple Mental Model

If none of the technical details have clicked yet, here is the simplest way to think about it.

Imagine you hire a contractor. You do not give them a building. You give them a blueprint and tell them what you need built. The contractor brings their own tools, their own workspace, their own expertise. You describe the outcome. They handle execution.

Managed Agents is the contractor.

You write down what the agent should do (AGENTS.md). You describe the tools it can use. You send a task. Google provides the workspace (the sandbox), gives the agent its tools, and runs the job. You get back the result.

The factory is not your problem. The product is.

Why This Matters

For developers: the time between "I have an idea for an agent" and "the agent is actually running" just collapsed. You are not scaffolding a sandbox, installing dependencies, managing environments, wiring tools. You are writing a markdown file and making an API call.

For teams: you can now build agents without needing a dedicated DevOps setup to support them. The infrastructure concern shifts from your team to Google Cloud.

For the broader picture: Managed Agents is not a standalone product. It is one level in a structured stack that lets teams start simple, move to more complexity when they need it, and have all their agents talk to each other via a shared protocol.

The hardest part of building agents was never the intelligence. It was everything around it.

Google just made everything around it someone else's problem.

Official documentation: Managed Agents API on Agent Platform

Google's announcement: Introducing Managed Agents in the Gemini API