Google Launches Managed Agents in Gemini API With Antigravity Preview

Google has rolled out Managed Agents in the Gemini API, a new capability that lets developers spin up a fully sandboxed AI agent with a single API call. Alongside the launch, Google introduced the Antigravity Agent — a general-purpose managed agent powered by Gemini 3.5 Flash that can write code, run commands, browse the web, and manage files inside an isolated Google-hosted Linux container.

The release moves Google into more direct competition with Anthropic's Claude Code and OpenAI's Assistants and Responses APIs, and is one of the most concrete agentic-AI primitives any frontier lab has shipped this year.

What Managed Agents actually do

A managed agent is a stateful agent that Google runs for you. You make one API call; Google provisions a fresh Linux sandbox, attaches Gemini 3.5 Flash as the reasoning engine, and gives the agent a working environment with persistent files.

The Antigravity Agent, the first managed agent Google is shipping, can do four things out of the box:

• Code execution: run Bash, Python, and Node.js. Install packages, run tests, build small applications.
• File management: read, write, edit, search, and list files. Files persist across interactions in the same agent session.
• Web access: Google Search and URL fetching to retrieve external data.
• Multi-step reasoning: chain the above into longer workflows without orchestration code.

The Antigravity Agent uses the same harness that powers Google's Antigravity IDE — the agent-first desktop product Google released earlier this year. That means developers can prototype an agent in the IDE and deploy the same harness via the API.

Why this matters: building production-quality AI agents has historically required developers to write thousands of lines of orchestration code — managing tool calls, sandboxing, state, retries, and security. Managed Agents collapses all of that into a single API endpoint.

How developers actually use it

Google is pushing a markdown-first developer experience. Instead of writing orchestration code, you define an agent's behavior in two files: AGENTS.md (the system prompt and high-level instructions) and SKILL.md (specialized capabilities the agent can invoke).

You then register the agent with the Managed Agents API and call it like any other Gemini endpoint. The agent runs in its own ephemeral Linux container that Google tears down when the session ends. This is the same pattern Anthropic uses for Claude Code's skills system, and it makes agent definitions portable across tools.

For enterprise customers, the same primitive is available through the Gemini Enterprise Agent Platform, also in preview. The enterprise version adds VPC-Service-Controls support, IAM-based access control, and audit logging.

Industry impact

The agent platform race now has three serious entrants: Anthropic (Claude Agent SDK and Claude Code), OpenAI (Responses API, Computer Use), and Google (Managed Agents, Antigravity). All three have converged on the same architectural pattern — a sandboxed environment with code execution, files, and web access — but the developer experience differs meaningfully.

Anthropic's pitch is depth: Claude Code is the most-adopted enterprise agent, and the Claude Agent SDK is the most flexible. OpenAI's pitch is reach: Responses API plugs directly into the ChatGPT ecosystem. Google's pitch with Managed Agents is simplicity and price: one API call, Gemini 3.5 Flash pricing (significantly cheaper than Opus or GPT-5.5), and zero infrastructure to manage.

For startups building agent products, the cost curve matters. Gemini 3.5 Flash is roughly an order of magnitude cheaper per token than frontier models like Claude Opus 4.8 or GPT-5.5, while benchmarks suggest it is competitive on agentic tasks. That makes Managed Agents an attractive default for high-volume agent workloads where Opus-class quality is overkill.

Expert perspectives

Reaction from the developer community has been notably positive. Several agent framework maintainers — including the team behind the Agno toolkit — have already published integrations for the Antigravity Agent, calling it "the cleanest managed-runtime API any frontier lab has shipped."

Skeptics have raised two concerns. The first is lock-in: agents defined in AGENTS.md and SKILL.md format are not yet portable across vendors, though Anthropic uses a near-identical convention. The second is observability — running agents inside a Google-hosted sandbox makes debugging harder than running them locally with full access to logs.

What's next

Three things to watch.

First, general availability. Managed Agents is currently in public preview. Google typically takes three to six months to move new APIs to GA, which would put the general release in late 2026.

Second, additional managed agents. Antigravity is the first, but Google has hinted at specialized managed agents for code review, data analysis, and customer support. Expect at least two more by year-end.

Third, integration with Workspace. Google has not yet announced a way to plug managed agents into Gmail, Docs, or Sheets, but it's the obvious next step — and the most direct competitive lever against Microsoft 365 Copilot and Anthropic's Claude-in-the-enterprise plays.

Bottom line

Google just made it dramatically easier and cheaper to build production AI agents. If you have been waiting for the agent platform race to produce something simple enough to start with, Managed Agents and the Antigravity Agent are the cleanest entry point yet. For developers already deep in Claude or OpenAI's tooling, the calculation is different — but the pricing alone is going to force a serious second look at Gemini for agent workloads.