Google DeepMind Unveils Interactions API and Managed Agents for Gemini

Google DeepMind has introduced two new primitives for building with Gemini: the Interactions API and Managed Agents. The Interactions API is a unified interface that uses server-side state, interaction IDs, and thought signatures to preserve context across multi-modal, multi-tool agent turns.

The Antigravity Managed Agent's sandboxes and credential-protecting proxy are also being launched as part of the new primitives. Ivan Leo, a member of the developer experience team at Google DeepMind, explained that as models get more capable, the scaffolding around them falls away, and the API surface has to change to match.

Leo presented several demos during his 17-minute conference talk, including one built by a colleague, Bamsi. The demo used a single photo with Nano Banana to generate variations of a person in different countries and locations. Each photo was a single interaction spawned from an initial API call, using the same interaction ID to carry through every call.

The new Interactions API replaces manual management of thought signatures, which were difficult to handle. Google found startups losing their cache over a single mistakenly added whitespace character due to this issue. The Interactions API addresses this by returning an interaction ID on the first turn, preserving context across multiple calls.