Google has expanded Antigravity and introduced managed agents in the Gemini API at I/O 2026. The 19 May announcements describe several ways to run an agent, from a desktop application and command line to a service that provides an isolated computing environment through an API.
The practical distinction is where the work runs and who maintains that environment. That affects how a project keeps its files, connects to other services and moves from an experiment into regular use.
A desktop, a terminal and a hosted option
Antigravity 2.0 is a standalone desktop application for coordinating agents. Google says it includes support for parallel work, dynamic subagents, scheduled tasks and connections across its development products. Antigravity CLI provides a command-line interface, while an SDK exposes the agent framework for developers who want to define behaviour and host it on their own infrastructure.
Managed Agents in the Gemini API take another approach. Google says a single API call can start an agent that reasons, uses tools and executes code inside an isolated Linux environment. The service uses the Antigravity agent framework with Gemini 3.5 Flash and is available through the Interactions API and Google AI Studio.
The environments retain files and state for follow-up interactions. That can help with work that needs several passes, because a later request can resume from the previous result. Persistence also means developers need to understand what is retained and which inputs belong in that environment before supplying business data.
AI Studio gets more routes into development
Google is also expanding AI Studio. The announcement lists native calls to relevant Google Workspace APIs, a path for exporting projects to local development in Antigravity, and Android application support. It includes publishing to a Google Play test track, which is different from a finished application being available to everyone in the store.
A separate AI Studio mobile application is offered for pre-registration during the announcement week. Pre-registration should not be treated as an immediate download or as proof that every workflow is available on a phone.
For a small team, these releases make the deployment question worth asking earlier. A prototype may work with a few supplied files, but regular use introduces permissions, costs and failure handling. Identify which account the agent uses, what it can modify and how a person can review the result before connecting production services.
The choice between a managed environment and your own infrastructure will depend partly on the controls and maintenance work you need. Start with a task whose success can be checked, such as processing a sample dataset or preparing a draft from a fixed document set, then expand access only as the workflow proves useful.
See AI Explorer for more on practical agent workflows.
Source: Google, “Building the agentic future: Developer highlights from I/O 2026”, 19 May 2026. Announcement image: Google.
