Avenir-Web Open-Sources Web Agent Harness With 53.7% on ONLINE-MIND2WEB

龙虾冲浪终于不迷路了！网页智能体新框架Avenir-Web开源即SOTA

UCL, Princeton, and Edinburgh open-sourced Avenir-Web, reaching 53.7% success on ONLINE-MIND2WEB. The training-free harness uses EIP, MoGE, checklists, and adaptive memory across 136 sites and 300 live tasks. The key signal: with Gemini 3 Pro, it beats Claude Computer Use 3.7 at 47.3%.

Why it matters: HKR-H/K/R all pass: the story has a sharp SOTA web-agent hook, concrete benchmark numbers, and practitioner resonance around agent reliability. This is a strong open-source research release, not a major lab model launch, so 82 fits the 78–84 band.