Researchers have built a pipeline that manufactures entire functional websites on demand — and the AI agents trained on those sites are getting meaningfully better at navigating the real web.
InfiniteWeb generates interconnected, multi-page web environments automatically for use in GUI agent training. Single-page generation with a large language model is a solved-enough problem, but a realistic site with navigation, forms, and linked content is harder to fake. The system tackles that gap through what the authors call unified specification and task-centric test-driven development, pairing a website seed with a reference design image to keep variety high. It also generates evaluators that can verify whether an agent actually completed a task, which enables the dense reward signals reinforcement learning depends on.
The training-data bottleneck is one of the quieter blockers in practical AI agent development. Agents that click, fill, and navigate real interfaces need exposure to a wide range of sites, but scraping the live web at scale raises legal and reproducibility problems. A synthetic-environment factory sidesteps both, and the benchmark numbers suggest the tradeoff is worthwhile — agents trained with InfiniteWeb showed notable gains on OSWorld and Online-Mind2Web, two standard GUI-agent benchmarks.
InfiniteWeb reportedly outperforms commercial coding agents at building realistic websites, which is either a sign of how good the system is or a reminder that most commercial coding tools were not designed for this specific, structure-heavy task — probably both.