Google DeepMind Simplifies Coding Agent Development with New Primitives

Google DeepMind has announced two new primitives that aim to simplify the development of coding agents. The Interactions API and Managed Agents are designed to collapse months of infrastructure work into individual API calls, according to Ivan Leo, a member of the developer experience team at Google DeepMind.

The Interactions API moves state management off the client's machine and onto Google's servers, reducing the risk of performance degradation due to manual errors. This is particularly important for newer models like Opus and Fable, which reason for extended periods and interact with an environment.

Leo demonstrated the power of the Interactions API by showing a multi-model pipeline built by his colleague Bamsi, where a single photo was fed to Nano Banana to generate variations of a person in different countries. The interaction ID was then used to generate video with the newly launched Omni model.

The second launch is Antigravity as a remote agent, which packages Google's coding harness into a persistent remote sandbox. This eliminates the need for developers to tune harnesses or provision infrastructure, and reduces the barrier to running up to 1,000 named agents.