Gemini API: Managed Agents and the Antigravity Agent in Public Preview

Google Launches Managed Agents in the Gemini API at I/O 2026

At Google I/O 2026 on May 19, 2026, Google introduced Managed Agents in the Gemini API — a hosted agent runtime that lets developers deploy a fully autonomous, stateful agent with a single API call, without any infrastructure management overhead.

What Managed Agents Are

Managed Agents are Google-hosted, isolated Linux sandbox environments that run an agentic loop on the developer's behalf. A single call to the Interactions API provisions the environment, starts the agent, and drives it through a plan-act-observe loop until the task is complete. Each agent can:

• Execute Bash, Python, and Node.js code with full output capture
• Read, write, edit, and search files within its sandbox
• Browse the web and fetch URL content
• Maintain persistent file state across multiple interactions within the same environment ID
• Handle long-running sessions via automatic context compaction

This removes the orchestration overhead that previously required developers to build and maintain custom agent harnesses, state managers, and sandboxed execution environments.

The Antigravity Agent

The primary Managed Agent available at launch is the Antigravity Agent, accessed with model ID antigravity-preview-05-2026. It is powered by Gemini 3.5 Flash and uses the same harness as Google's Antigravity IDE — meaning the hosted API infrastructure is the same system running Google's own internal agentic workflows.

Access follows a straightforward pattern: the developer specifies an agent_id, an input prompt, and an optional environment parameter (either a fresh "remote" instance for an ephemeral session, or an existing environment ID for session continuity). The agent autonomously plans and executes until the task resolves.

Building Custom Managed Agents

Beyond the Antigravity Agent, developers can register and deploy their own custom managed agents. Instructions and capabilities are defined in markdown files — AGENTS.md for agent-level instructions and SKILL.md for reusable capability definitions. These files can be version-controlled and deployed as a managed agent via the Gemini API, making the entire agent configuration a standard software artifact rather than live runtime state.

Google Cloud Storage and repository mounts are also supported, enabling agents to operate directly on a team's codebase or data assets without manual file transfer.

Pricing During Preview

Environment compute — CPU, memory, and sandbox execution — is not billed during the public preview period. Developers pay only for standard Gemini 3.5 Flash token consumption. Based on usage patterns documented in the preview, research tasks typically consume $0.30–$1.00 in token charges; complex data processing pipelines run $0.70–$3.25. This makes the preview period essentially free for experimentation beyond token costs.

Availability

Managed Agents are in public preview via the Gemini API and Google AI Studio Playground. Enterprise deployments are supported through the Gemini Enterprise Agent Platform on Google Cloud, with full IAM integration and Google Cloud project connectivity.

What This Means for Developers

For Gemini API and Gemini CLI users, Managed Agents remove the most significant operational barrier to shipping agentic applications in production: the need to provision, secure, and maintain an execution environment. Tasks that previously required a custom agent harness, a sandboxed execution layer, and persistent state logic can now be expressed as a single API call with a markdown-defined agent configuration. The free compute tier during preview makes exploration essentially zero-cost beyond standard token usage.