For browser agents to become ubiquitous, they need to be faster, cheaper, and more reliable. Unfortunately, the most popular solutions today are either really bad (browser-use scores 31% on BU Bench v1 with GPT 5.6 Luna) or token inefficient (BrowserCore scores 78% on BU Bench v1, but consumes 100M+ tokens. So I decided to build a dedicated browser agent harness called Browser Agent, that outperforms BrowserCode on BU Bench v1: Browser Agent: - Success Rate: 88% - Cost: $5.37 for 100 tasks - Total aggregate duration: 32,694 seconds Browser Code: - Success Rate: 78% - Cost: $8.34 - Total aggregate duration: 47,970 seconds Browser Agent produces higher quality trajectories, while using up to 91% fewer tokens. This has two benefits: it reduces the risk of context rot and substantially lowers VRAM and compute requirements to fine-tune small models on the harness' trajectories. Therefore, this harness is designed for: 1) Users of browser agents who want to do more with less 2) Teams training small models to reach SOTA performance at web browsing

Browser Agent

An efficient harness to automate any task on the web

An efficient harness to automate any task on the web