Google's Managed Agents API lets you spin up a full agentic sandbox with one API call
Google's new Managed Agents API provisions a Gemini-powered agent inside an isolated Linux sandbox in a single API call — no infrastructure required.
Building a production-grade AI agent used to mean stitching together a model, a code interpreter, a file system, a browser tool, and some kind of sandboxed execution environment — before writing a single line of actual business logic. Google announced at I/O 2026 that it is collapsing all of that into a single API call.
What Google launched
Managed Agents in the Gemini API is a new primitive that provisions a fully operational agent inside a Google-hosted Linux sandbox the moment you call the Interactions API. The agent can reason and plan, write and execute code (Python, Node.js, Bash), manage files, install packages, and browse the web — all inside an isolated container that Google manages for you.
The agent powering this is called the Antigravity agent (antigravity-preview-05-2026), built on Gemini 3.5 Flash. Antigravity itself is Google’s agent-first development platform, relaunched as version 2.0 alongside this announcement. The Managed Agents feature is the hosted, API-accessible version of that platform.
The quickstart shows three parameters: the agent model name, environment="remote" to provision a fresh sandbox, and an input describing what you want the agent to do. That is your entire setup.
What the agent actually does inside the sandbox
Once the sandbox spins up, the Antigravity agent runs a full autonomous loop. It breaks down your task, writes code to solve it, executes that code, reads the output, and continues reasoning until the task is complete. If it needs to fetch live data from the web, it does that too.
One detail worth highlighting: the sandbox is stateful. Most agent APIs today are fire-and-forget — the agent runs, the environment is destroyed, and nothing persists. With Managed Agents, you can resume the same environment across multiple API calls. Files you wrote in call one are still there in call two. Processes can stay running between interactions. This makes it practical for longer workflows that unfold over time, rather than single-shot tasks.
Security is locked down by default. Out of the box, the sandbox has no access to external networks, credentials, or production systems. Any connectivity beyond the container has to be explicitly configured. If you need the agent to write results back to Cloud Storage, you set that up deliberately — it does not happen automatically.
Gemini 3.5 Flash as the engine
The Antigravity agent runs on Gemini 3.5 Flash, which Google also launched at I/O 2026. The numbers are worth noting: it outperforms Gemini 3.1 Pro on agentic and coding benchmarks (Terminal-Bench 2.1: 76.2%, MCP Atlas: 83.6%) while running four times faster than comparable frontier models, often at less than half the cost. For the kind of multi-step reasoning loops that agents run, speed matters — slower models create noticeable lag across many tool calls.
Customising the agent
If the general-purpose agent does not fit your use case, you can extend it without writing orchestration code. Google lets you define custom instructions and skills in markdown files (AGENTS.md, SKILL.md) and register them as a named agent. You can also use the Antigravity SDK if you want full programmatic control and the option to self-host on your own infrastructure rather than using Google’s managed environment.
What this means for you
If you have tried to build an agent that goes beyond a simple chatbot, you have probably spent a meaningful amount of time on plumbing that has nothing to do with your actual product. Setting up sandboxed code execution, handling file state, wiring in a browser tool, managing the execution loop — none of that is differentiated work, but it is all time-consuming and easy to get wrong.
Managed Agents removes that entirely for the common case. You describe a task, the API provisions an environment, and Gemini handles the rest. For prototyping, that is a significant acceleration. For production use, the security defaults (isolated sandbox, no implicit external access, Google-managed infrastructure with enterprise data governance) mean you are not starting from scratch on compliance either.
The preview pricing is straightforward: you pay for Gemini model tokens and any tools used. Sandbox compute is free during the preview period.
Current limitations to know about
This is a preview, and the rough edges are documented. Structured output is not supported yet. Several generation config parameters (temperature, top_p, top_k) return a 400 error if passed. Tools including file_search, computer_use, google_maps, and function_calling are not available in this release. If any of those are requirements for your use case, you will need to wait for a later version or build around them.
The broader picture
Managed Agents is one piece of a larger push from Google at I/O 2026. Alongside the Antigravity 2.0 desktop app and native Android vibe coding in Google AI Studio, the through-line is consistent: Google is building Gemini as an agent platform, not just a model API. The shift from “call the model, get text back” to “call the API, get a task done” is exactly what Managed Agents represents.
For developers who have been watching agentic AI from the sidelines waiting for the infrastructure story to mature, this is a reasonable moment to start experimenting.