Google opens Managed Agents preview in the Gemini API

Google announced Managed Agents in the Gemini API in Public Preview at I/O 2026 on May 20, exposing the same Antigravity-powered agent harness that backs Google's own developer tools through a standalone API. The pitch: one API call provisions an isolated, ephemeral Linux environment where Gemini 3.5 Flash can reason, plan, call tools, execute code, manage files, and browse the web. State persists across follow-up calls, so multi-turn sessions resume with files intact.

What Managed Agents actually does. Per Google's developer blog and Gemini API documentation, calling Managed Agents through the Interactions API hands back an agent session running in a managed sandbox. Within that session, the agent can use a defined toolset, write and execute code in supported languages, read and write files, and access the web through a managed browser. Each interaction creates an environment that the developer can resume in a follow-up call, with all files and state intact — enabling multi-turn workflows that previously required the developer to manage their own sandbox infrastructure. Developers can customize the agent inline (instructions, skills, data) at interaction time, or save the configuration as a named managed agent invokable by ID.

Built on the Antigravity harness. Google framed Managed Agents as the same agent harness that powers Antigravity (Google's developer-facing AI platform, updated earlier in the week at I/O). The underlying model is Gemini 3.5 Flash — the agentic-first model that Google made the default at I/O. This means the agent runtime, the tool-use protocol, and the model are all owned by Google end-to-end. For developers, the trade-off is convenience vs. lock-in: less plumbing to maintain, but the runtime is not portable to other models.

Preview status and what's not yet stated. Managed Agents is in Public Preview through the Interactions API and Google AI Studio. Public Preview means generally available to developers but subject to change before GA, with no formal SLA. Pricing for preview usage was not detailed in keynote materials, nor were rate-limit specifics — Google's documentation lists feature behaviour, not preview limits. Developers planning production workloads on Managed Agents should treat current behaviour as subject to change and verify pricing before scaling.

Why this matters for AI tool users. For developers and small teams building AI applications, Managed Agents lowers the floor for "I want my own agent" projects. The pattern was previously: stand up your own sandbox, manage state between tool calls, wire up code execution and file I/O. The new pattern is one API call. This puts Google's agent infrastructure closer to where OpenAI's Assistants API and Anthropic's Claude Code already sit — competitive parity on the infrastructure layer, where the differentiation is now model quality, pricing, and ecosystem fit. See our best AI coding tools roundup for the agent-builder landscape, and our Emergent vs Bolt.new comparison for two AI app builders that take different approaches to managed runtime.

What to do this week. If you build agent workflows today on Claude or OpenAI APIs, evaluate Managed Agents against your existing setup — but on its merits, not the keynote framing. Run one realistic workflow (multi-step tool use, file I/O, web browsing) through Managed Agents and measure latency, output quality, and total cost against your current provider. Don't commit production traffic until preview-status pricing and rate limits are formally published. If you're earlier in the build cycle, our Emergent review covers a no-code agentic app builder that uses similar infrastructure patterns, and the automation/" class="text-blue-300 hover:text-blue-200 underline decoration-blue-400/40 underline-offset-2">workflow automation roundup compares the non-agentic alternatives (Zapier, Make, n8n) for problems that don't need a full agent.