创业公司Hark近日预览了其浏览器使用Agent产品Handoff，该工具能够通过访问网站的结构和视觉数据来自主完成订餐、预订、购物等各类在线任务，无需借助官方API接口。[1]Handoff已被演示可与Target、Walmart、OpenTable和LinkedIn等多个主流网站进行交互。[1]该公司CEO Brett Adcock在演示中展示了该Agent能够理解模糊表述（如"部分花商精选"）来完成订花任务的能力。[1]
Hark声称其模型设计的独特之处在于专注于预测下一步行动而非预测下一个词元，这使其性能更快、成本更低，相比GPT-4.5和Opus等竞品具有优势。[1]该公司已于5月完成Series A融资，金额达7亿美元。[1]Hark目前已开放Handoff的等待名单，计划在夏末正式发布该产品，并在今年晚些时候进行预训练。[1]
Startup Hark has unveiled Handoff, a browser use agent designed to automate web-based tasks such as food ordering, booking reservations, and shopping without requiring official APIs from target websites.[1] The agent operates by accessing website structure and visual data to complete user requests, enabling interaction with platforms including Target, Walmart, OpenTable, and LinkedIn.[1]
According to CEO Brett Adcock, the model distinguishes itself by focusing on predicting the next action rather than the next token, delivering faster performance and lower costs compared to competitors like GPT-4.5 and Opus.[1] During a demonstration, Adcock showcased Handoff's ability to handle ambiguous language—such as "florist's choice" selections—to successfully complete flower delivery orders.[1]
Hark, which raised $700 million in Series A funding in May, plans to launch Handoff by late summer with a waitlist currently available for interested users.[1] The company also intends to conduct pretraining work on its model later this year.[1]