Summary
Philipp Schmid from Google DeepMind presents a live-coded progression showing how agent development has evolved from raw Python loops to YAML configuration files and bare file systems — each iteration deleting code rather than adding it. The central thesis: files are replacing Python as the primary medium for defining and running agents.
The talk introduces the Gemini Interactions API, Google’s new unified interface for calling both Gemini models and hosted agents. Unlike traditional turn-based conversation APIs, the Interactions API structures execution as a flat timeline of typed steps — user input, reasoning, function calls, function results — removing the need to abuse the user role for returning tool data. Schmid live-demos a GitHub PR review agent built three ways: first as a raw Python loop with hand-authored JSON schemas, then using the ADK agent framework, and finally as a near-codeless configuration pointing at a bash tool and a file system.
The second half announces that the anti-gravity agent — the same harness powering Google’s anti-gravity IDE — is now available through the Gemini API as a general-purpose agent. A new `environment` parameter lets developers attach isolated cloud sandboxes to agent calls, pre-loaded from GitHub repositories or GCS buckets. A network proxy injects credentials on behalf of the agent so API tokens are never exposed inside the sandbox. Schmid also introduces a persistent agents API that lets developers register custom agent configurations — system instructions, tools, environments — and reference them by ID, the same way they reference Gemini model IDs.
📺 Source: AI Engineer · Published September 14, 2026
🏷️ Format: Keynote Launch
