Google has launched managed agents in the Gemini API, allowing developers to make a single API call to spin up an autonomous agent running in a sandboxed remote Linux environment. The agent can write and execute code, run bash commands, and use tools without touching production servers. Built on the anti-gravity agent harness (which also powers the anti-gravity IDE), developers can customize agents with their own system instructions and skills defined via markdown files. The new Interactions API underpins this, replacing the older generate-content model with an agent-first data model that supports multi-step workflows, sub-agents, tool calls, and streaming steps. Deep Research was the first managed agent launched under this API. A demo showed a radio production app that reads Hacker News, picks opposing views, writes a script, generates audio via Gemini TTS and Lyria, and produces a 3-minute radio show — all from a single API call. Developers can get started at the Gemini API docs, AI Studio, or agent.new.