WebAnything Automates Web Data Collection for AI Agents

A team of researchers has built a system called WebAnything that turns live websites into training grounds for AI agents, no human engineering required per site. The pitch is straightforward: training agents to navigate the web reliably has always run into a bottleneck. Either someone manually mirrors a site, which takes custom work for every URL, or a model generates a fake version of a site, which tends to drift from how the real thing actually behaves. A third option, recording and replaying site interactions, locks researchers into whatever was captured the first time around, limiting how much an agent can actually explore.

WebAnything tries to sidestep all three problems. It's an automated pipeline that pulls real websites into reusable, resettable environments, built from actual URLs rather than synthetic approximations or static recordings. The researchers tested the resulting environments against a benchmark called REAL, along with Odysseys-Easy and WebArena-Verified, to see whether agents trained on this data actually get better at completing tasks. They also fine-tuned a Qwen3.5-9B model on interaction data pulled from the system to check real-world performance gains.