An Interaction Is All You Need — Ivan Leo, Google DeepMind

Ivan Leo, a developer experience engineer at Google DeepMind, gave this talk at the AI Engineer World's Fair 2026 in San Francisco. It introduces two products, the Gemini Interactions API and Managed Agents (0:12).

He opens with history. From 0:27 he traces the move from single completions to agents, by way of function calling (1:02) and models that reason and act (1:32). At 2:02 he offers a definition of an agent, and at 3:02 he points viewers to Google AI Studio for trying the models.

The case for a new API starts at 4:07. The description puts it simply: models became agents and the APIs did not keep up. At 5:31 he explains server-side state, where interaction IDs replace the job of passing thought signatures back by hand. A demo at 6:21 chains image and video generation, and 7:01 walks through the code behind it. Three shorter chapters follow: strongly typed outputs (7:36), mixing built-in and custom tools (8:21) and the new steps data model (9:01).

Managed Agents take the second half, from 9:46. They run the Antigravity agent in a persistent remote sandbox. A demo at 10:36 analyzes a GitHub repo. Sandboxes persist through environment IDs (11:06), and sources can come from Google Cloud Storage, GitHub or inline files (11:41). The chapter at 12:16 is a 2M-token repo analysis, and 12:56 covers running the same agent locally and in the cloud. A proxy that injects credentials comes up at 13:36, then named agents at 14:26.

The last stretch covers the open-source Gemini API CLI (15:26) and a skill for migrating code to the Interactions API (15:56), with a wrap-up at 16:36.