BrowserForge: Scaling Web Agent Training Data via Parallel Browser Sandboxes

Forum topic · 小凯 · 2026-08-27

Summary

BrowserForge is a research framework (arXiv:2608.24848, authors Fei Tang, Huawen Shen, Zhiqiong Lu, Zhengxi Lu, Pengyuan Lyu, et al.) that addresses the data bottleneck in training web agents that act directly from rendered pixels. While pixel-based agents avoid the fragility and high token costs of reading HTML or accessibility trees, existing public datasets contain only a few thousand trajectories from a narrow, fixed set of websites, and recent automated synthesis pipelines remain tied to predefined site lists or tutorial sources. BrowserForge generates web interaction data at scale by driving many browser sandboxes in parallel over the open web. It couples three components: (1) an open-web sourcing stage exposing agents to hundreds of thousands of real, publicly reachable websites; (2) a sandbox cluster manager scheduling hundreds of concurrent browsers with high utilization; and (3) a proposer-solver dual-agent loop that converts raw pages into executable tasks and collects verified trajectories. The framework substantially expands the diversity of websites seen during training, tackling the long-standing problem of producing high-quality web interaction trajectories at scale.

Overview

Field: NLP
Authors: Fei Tang, Huawen Shen, Zhiqiong Lu, Zhengxi Lu, Pengyuan Lyu, et al.
Published: 2026-08-25
arXiv: 2608.24848

Key Points

Problem: Web agents that act from rendered pixels avoid the fragility and heavy token cost of reading a page's HTML or accessibility tree, but training them depends on large amounts of high-quality interaction trajectories — and producing such data at scale remains an open problem.
Limitations of existing data: Public datasets typically contain only a few thousand trajectories drawn from a fixed and narrow set of websites. Even recent automated synthesis pipelines stay bound to predefined site lists or tutorial sources, so the number of distinct websites the agent ever sees barely grows.
Proposed solution: BrowserForge, a framework that generates web interaction data at scale by driving many browser sandboxes in parallel over the open web.

Framework Components

BrowserForge couples three components:

1. Open-web sourcing stage — exposes the agent to hundreds of thousands of real, publicly reachable websites, dramatically increasing site diversity beyond curated lists. 2. Sandbox cluster manager — schedules hundreds of concurrent browsers while achieving high utilization across the cluster. 3. Proposer–solver dual-agent loop — transforms raw pages into executable tasks, then collects verified trajectories for training.

Significance

By decoupling data generation from predefined site lists, BrowserForge offers a scalable path to producing diverse, verified web interaction trajectories for training pixel-based web agents.

--- *Auto-collected on 2026-08-27.*