AI BEACON #03 — World models go live, agents move into the browser

Capability gains are showing up as execution, not just better text — models are writing code, using tools, and operating inside default interfaces. Google pushed agentic vision into Gemini 3 Flash and started rolling out an Auto Browse agent in Chrome, while Project Genie moved from concept to a live Ultra-subscriber product in the U.S. At the same time, frontier pressure is being expressed through tool-using reasoning (Qwen3-Max-Thinking) and open-source scale (Kimi K2.5). The implication is operational: autonomy will scale through where agents are embedded — browsers, data clouds, writing surfaces — and the organizations that control those surfaces will set the constraints.

TL;DR

• Browsers and workflow surfaces are becoming agent runtimes → permissions and UI control matter more.
• Tool-using reasoning is the competitive baseline → execution loops become the evaluation story.
• Open-source scale plus falling integration friction → multi-model strategies get easier to justify.