An Interaction Is All You Need — Ivan Leo, Google DeepMind
Source
youtube.com
Author
AI Engineer
Date
Context
Google DeepMind's Interactions API lets an application continue an AI conversation by referring to its previous interaction. The server retrieves the stored history, so the application does not have to resend the entire conversation on each turn. The documentation says this server-side state management is optional.
Each interaction also records execution steps, including model thoughts, tool calls, their results and the model's output. That gives developers a structured account of a task that involves several actions. This is the state-management change described in the conference talk, with conversation continuity and execution history handled as explicit API resources.
Key takeaways · AI-distilled
The Interactions API keeps conversation state server-side behind interaction IDs, so developers no longer hand-manage Gemini's thought signatures between turns.
One context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → can chain image, video and audio generation, return strongly typed multimodal outputs, and mix built-in tools with custom functions, with results organized in a new 'steps' data model.
Managed Agents run Google's Antigravity agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → in persistent remote sandboxes addressed by environment IDs, loading sources from GCS, GitHub or inline files. Leo demos a 2M-token repository analysis.
Managed Agents also add a credential-injecting proxy and named agents. Leo shows the same AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → running locally and in the cloud, plus an open-source Gemini API CLI and a agent skillA reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.Full definition → for migrating to the Interactions API.
Terms in this piece · Glossary
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.