BrowserForge: Scaling Web Episode via Parallel Browser Sandboxes

TL;DR

BrowserForge generates large-scale web interaction data by running hundreds of browser sandboxes in parallel across the open web. It uses a dual-agent system to create executable tasks from real pages and produces a corpus of 203,238 diverse, verified trajectories, enabling better training of pixel-based web agents.

Why It Matters

Scaling high-quality web interaction data is critical for training robust, generalizable web agents. BrowserForge enables this at unprecedented scale and diversity, directly improving agent performance on real-world benchmarks without relying on synthetic or limited site sources.

Key Findings

• BrowserForge scales data collection by orchestrating hundreds of concurrent browser sandboxes across hundreds of thousands of real websites.
• The Proposer-Solver loop generates executable tasks from raw web pages and collects verified trajectories, with a rule-plus-model pipeline cleaning and standardizing the reasoning.
• Training a multimodal model on BrowserForge's corpus improves success rates on live and static web benchmarks, with gains increasing as dataset size grows.