https://video.ut0pia.org/videos/watch/a1aeec4b-2edf-4b9a-a87e-577a1520ce2a Every browser agent improvement in the last year has mostly been a model upgrade: better vision, longer context, smarter planning. And they still fail on basic workflows. Our claim is that the main bottleneck is not the model. It is the interface we give the model to the browser. Three things matter more: what the model sees, what it can do, and what it learns from. We built a browser-agent runtime around all three: a compact page representation instead of a raw dump, fast actions with stable handles instead of one click per call, and step-by-step feedback instead of pass/fail at the end. In our early runs, changing that interface alone was enough to take the same model from confusion to correct multi-step execution on hostile pages. This talk is the thesis, the evidence, where it still breaks, and why better browser state is a bigger lever than just swapping in a better model. Speakers: Kushan Raj (ARK): Kushan Raj is a Founding ML Engineer at Sarvam AI, where he built the real-time voice AI stack that now powers 2M+ daily calls across 10+ Indian languages LinkedIn: https://www.linkedin.com/in/kushanraj/ Mon, 29 Jun 2026 15:48:07 GMT https://validator.w3.org/feed/docs/rss2.html PeerTube - https://video.ut0pia.org https://video.ut0pia.org/lazy-static/avatars/0287a09a-aae7-4840-9843-b416426e7046.webp https://video.ut0pia.org/videos/watch/a1aeec4b-2edf-4b9a-a87e-577a1520ce2a All rights reserved, unless otherwise specified in the terms specified at https://video.ut0pia.org/about and potential licenses granted by each content's rightholder.