Summary
Kushan Raj, a founding engineer at ARK (formerly at Seraphim), argues that browser agents are failing not because models are too weak, but because the environment they operate in is too poor. In a live demo using the browser challenge benchmark, he shows how Claude gets stuck on trivially simple tasks — spending two minutes to download an Aadhaar document or failing entirely to pick a date on a booking site — due to limited page visibility and no structured action feedback.
Raj’s solution is a compressed page representation that converts a full DOM (typically 20,000+ tokens) into roughly 1,800 tokens of clean markdown that covers the entire page at once — cheaper than a screenshot at 1,100 tokens, but with dramatically more context. The system also tracks page state changes between actions, telling the agent what elements appeared, disappeared, or blocked a click. This gives the model enough grounding to plan long, reliable action sequences rather than looping on confused screenshots.
Using this infrastructure with a cheaper model than Claude, Raj’s agent completes the same tasks in seconds. He plans to open-source the project and potentially expose it as an API where developers pass a URL and intent to receive automated execution results. The talk is a practical argument that browser-agent adoption is an infra problem, not a model problem.
📺 Source: AI Engineer · Published June 28, 2026
🏷️ Format: Showcase
