Google launches Gemini Spark managed sandboxes and a cloud agent for background tasks
- Google on May 19 launched Managed Agents in the Gemini API and introduced Gemini Spark, a cloud-based agent that can keep running tasks offline. - Google said a single API call can start an agent in an isolated Linux environment, while Spark will reach U.S. AI Ultra subscribers next week. - Google’s developer posts and Gemini app rollout pages list the next steps, including trusted-tester access and a U.S. beta expansion.
Google used its May 19 I/O 2026 announcements to add two missing pieces for developers trying to move AI agents from demos into production: a managed runtime and an always-on cloud agent. The company introduced Managed Agents in the Gemini API, which let developers start an agent that can reason, use tools and execute code inside an isolated Linux environment, and it unveiled Gemini Spark, a personal agent designed to keep working in the cloud even when a user’s device is offline. Google described both products as part of a broader push toward “action” rather than one-shot responses. The announcements were spread across Google’s developer, model and Gemini app posts published during I/O. ### What exactly did Google launch on May 19? Google said on May 19 that the Gemini API now supports Managed Agents, a product that lets developers “spin up an agent” with a single call. In Google’s description, that agent can reason, call tools and execute code in an isolated, ephemeral Linux environment, with agent behavior defined in files such as `AGENTS.md` and `SKILL.md`. (blog.google) Google also introduced Gemini Spark in consumer-facing materials the same day. The company called Spark a “24/7 personal AI agent” that runs on Gemini 3.5 Flash and is designed to proactively manage tasks under a user’s direction. Google said Spark can continue operating in the cloud rather than depending on a laptop or phone staying on. ### Why do managed sandboxes matter for agent builders? (blog.google) Google’s developer post said Managed Agents run in “secure cloud sandboxes,” and the launch page described the runtime as an isolated Linux environment. That matters because many agent workflows need somewhere to execute code, inspect files, call tools and recover from failures without exposing a developer’s own machine or requiring teams to build container orchestration from scratch. (blog.google) The company’s framing also puts more of the agent stack under Google’s control. Google said developers can define custom agents with their own instructions, skills and data, while the managed runtime handles execution inside Google’s hosted environment. The official posts do not promise to solve every production concern, but they do move setup away from local scripts and toward a service model. (blog.google) ### How is Gemini Spark different from a chatbot tab left open in a browser? Gemini Spark was presented by Google as a background agent, not just a conversation interface. The Gemini app post said Spark is built to “proactively manage tasks” and help users navigate their digital lives, while the Gemini 3.5 post said it “runs 24/7” and can take action on a user’s behalf under that user’s direction. (blog.google) That means Spark is closer to an asynchronous worker than to a chat session that stops when the window closes. Google’s materials indicate the service is meant to keep operating in the cloud, and the company said the macOS app will integrate Spark so it can also operate on a local machine. ### Where does this leave the work developers still have to do? (blog.google) Google’s own launch language points to orchestration work rather than prompt-writing alone. Managed Agents can reason, use tools and execute code, but developers still need to define instructions, skills and data boundaries, and they still need to decide how agents should behave across multi-step tasks. (blog.google) The practical issues are the familiar ones for long-running software: permissions, retries, visibility into what happened, and when to keep work synchronous versus handing it to a background process. Google did not spell out a full operations checklist in the launch posts, but the products it introduced are explicitly built around hosted execution and persistent task handling rather than prompt-response loops alone. That is an inference from Google’s product descriptions and deployment model. (blog.google) ### Who gets access first, and when? Google said on May 19 that Gemini Spark was rolling out first to trusted testers. In the Gemini 3.5 launch post, the company said it plans to bring the Spark beta to Google AI Ultra subscribers in the United States “next week.” Google’s next public markers are on its own product pages. Developers can already access Managed Agents through the Gemini API materials, while Spark’s next expansion point is the U.S. (blog.google) AI Ultra beta rollout Google said would follow the initial tester release. (blog.google)