News
What Are Gemini Managed Agents? Google’s Hosted AI Agents, Explained
Gemini Managed Agents let one Gemini API call start an AI agent in its own Linux sandbox. What they are, what they cost, and the new AI Studio hub on the way.
"Gemini Managed Agents" keeps showing up in developer feeds, and on October 25, 2026, a report said Google is building a whole new section of AI Studio just for them. So what are they, and why does Google want a dedicated home for them?
The short version: a managed agent is an AI helper that Google runs for you, on Google’s computers, from a single request. Below is what that means in plain words, what it costs today, and what the new hub would add.
What are Gemini Managed Agents?
Most AI calls work like a question and an answer. You send text, the model sends text back, and you do the rest. A managed agent works more like hiring a helper with a desk and a computer. You give it a goal, and it keeps going on its own: it makes a plan, searches the web, writes and runs code, saves files, and returns finished work.
Google announced the feature in a blog post by Ali Çevik and Philipp Schmid of Google DeepMind on May 19, 2026, during I/O week. The key idea is that Google hosts everything. Developers no longer have to build their own sandbox, tool loop and file storage. Google opened up the same agent harness and infrastructure it uses for its own agents. Its first managed agent was Deep Research, launched in December 2025.
How do Gemini Managed Agents work?
Every managed agent is built on the Antigravity agent, the same agent brand Google uses in its Antigravity coding tools. It launched on Gemini 3.5 Flash. Since a July 28, 2026 update, it runs on Gemini 3.6 Flash by default, and developers can switch to Gemini 3.5 Flash or the cheaper Gemini 3.5 Flash-Lite.
Here is what happens when a developer sends a request:
- A sandbox starts. The agent gets a secure, short-lived Linux machine. Google’s docs list built-in tools for Google Search, reading web pages, running Bash and Python code, and working with files.
- The agent works in steps. It reasons, calls tools and checks its own results until the job is done.
- State is kept. Follow-up requests can resume the same sandbox with its files intact. Unused sandboxes expire after 7 days.
- Network access is limited. Outside internet access is restricted by default through an allowlist that the developer controls.
Developers can make their own custom agents by adding instructions and skills as plain Markdown files, named AGENTS.md and SKILL.md. Google’s example is a data analyst agent that reads quarterly revenue numbers and builds a slide deck. If skills sound familiar, Google uses the same idea for regular users in the Gemini app, which we explain in What are Gemini skills?
That is the same file format coding agents like Codex and Claude Code read. Here is what AGENTS.md is and why it matters for security.
Gemini Managed Agents timeline
| Date | What changed |
|---|---|
| December 2025 | Deep Research becomes Google’s first managed agent |
| May 19, 2026 | Managed Agents open in preview in the Gemini API and the AI Studio Playground, powered by Antigravity on Gemini 3.5 Flash |
| July 7, 2026 | Background tasks for long jobs, remote MCP servers, function calling and credential refresh, announced by Google’s Logan Kilpatrick |
| July 28, 2026 | Gemini 3.6 Flash becomes the default, plus free tier access, token budgets, scheduled triggers, safety hooks and an Environments API |
| October 25, 2026 | TestingCatalog reports a dedicated Managed Agents hub in AI Studio is in development (not announced) |
Are Gemini Managed Agents free?
Partly. Since July 28, managed agents work on free tier projects, so a developer can test them with an API key from a project without billing turned on. On paid projects, Google’s pricing page says all model use is charged at standard Gemini list rates, and the sandbox itself (CPU, memory, running code) is not billed during the preview.
The catch is that agents can use a lot of tokens, because they loop through many steps on their own. Google warns about this directly and added a max_total_tokens budget. When an agent hits its budget, it stops and marks the job as incomplete, and the developer can resume it with a new budget. Google has not said what sandbox compute will cost after the preview ends.
The new Managed Agents hub in AI Studio
Today, developers can try managed agents without code by opening the Playground in Google AI Studio and flipping the toggle to Agents. Google offers open-source starter templates in the google-gemini/gemini-managed-agents-templates repository on GitHub.
According to TestingCatalog, Google is now building a separate Managed Agents section with five areas: Agents, Environments, Credentials, Triggers and Runs. That would put setup, connected accounts, schedules and run history in one place, instead of spread across code. This is a leak based on an unreleased interface, and Google has not confirmed a rollout date. TestingCatalog also suggested the hub could use Google Cloud’s new Agent Compute Engine, but that link is not confirmed either.
Managed agent vs a regular Gemini API call
| Regular Gemini API call | Gemini Managed Agent | |
|---|---|---|
| What you get back | One reply | Finished work: files, code results, reports |
| Runs code? | Only with extra setup | Yes, in its own Linux sandbox |
| Remembers files between calls? | No | Yes, until the sandbox expires |
| Can run on a schedule? | No | Yes, with triggers |
| Who builds the agent loop? | The developer |
Businesses can also use managed agents through the Gemini Enterprise Agent Platform, which is in preview too.
What this means for you
Gemini Managed Agents are a tool for developers. To use one, you need an API key, some code or the AI Studio Playground, and a plan for token costs. That is a big step forward for people building apps, because the hard part of an agent (the sandbox, the tool loop, the memory) now comes ready-made.
It also shows where AI is going. The biggest companies are moving past chatbots that answer questions toward agentic AI that does the work. If you want to see that loop step by step, read How AI agents actually work.
If you are not a developer, you do not need to set up any of this to get the benefit. Agentic AI for iPhone and iPad gives you a team of ready-made expert agents that plan and finish tasks like research, writing and trip planning. There is no account and no login, the first 3 tasks are free, and you can put up to 4 agents on one job.
Quick answers
What are Gemini Managed Agents? A Gemini API feature, in preview since May 19, 2026, that starts a Google-hosted AI agent from one API call. The agent plans, uses tools, browses the web and runs code in its own temporary Linux sandbox.
What model do Gemini Managed Agents use? They run on Google’s Antigravity agent. Since July 28, 2026, it uses Gemini 3.6 Flash by default, with Gemini 3.5 Flash and Gemini 3.5 Flash-Lite as options.
How much do Gemini Managed Agents cost? Model use is billed at standard Gemini API rates, and sandbox compute is not billed during the preview. Free tier projects can use them too. Developers can cap spending with a token budget.
How do I create agents in Gemini? Developers can open Google AI Studio, go to the Playground, switch the toggle to Agents and start from a template, or create agents in code with the Agents API and run them through the Interactions API.
Do Gemini Managed Agents support MCP? Yes. A July 2026 update added remote MCP server support, along with background tasks and function calling. For a related idea on the Claude side, see What are MCP Tasks?
Is the AI Studio Managed Agents hub out? No. As of October 25, 2026, it has only been spotted in development by TestingCatalog. Google has not announced it.
Sources: Google: Introducing Managed Agents in the Gemini API · Google: Managed Agents, 3.6 Flash, hooks and more · Gemini API docs: Agents in AI Studio · Gemini API pricing · TestingCatalog · Gigazine · RuntimeWire
