Hark previews its browser use agent for completing tasks

Hark, a venture-backed startup that raised $700 million in Series A funding in May, has unveiled its innovative browser-based agent, Hark Handoff. This new tool can seamlessly navigate websites lacking official APIs, including e-commerce giants like Target and Walmart, as well as platforms such as OpenTable and LinkedIn. The Handoff agent intelligently analyzes the structure and visual data of web pages to determine if it needs to click buttons or input data manually.

Users can issue commands to the assistant, which can handle a wide range of tasks, from ordering food and coffee to booking travel tickets, filing tax returns, shopping for essentials, reserving restaurant tables, and conducting research on various platforms. In a demo video, Hark CEO Brett Adcock showcased the assistant's ability to build a flower bouquet based on specified criteria, including less precise terms like "some of the florist's choice."

The demo, while not fully representative of the product's capabilities, hints at the agent's potential. Hark is currently employing a post-trained model for this release, with plans to pre-train the model later this year to enhance its data pipeline, training infrastructure, and performance. Unlike large language models that predict the next token, Hark's model is designed to predict the next action, which could involve either a clock or a keyboard input at a specific location.

The startup contends that its Handoff agent outperforms competitors in terms of speed and affordability, offering a more cost-effective solution compared to other models like GPT 5.5 and Opus 4.8. Hark has opened a waitlist for its platform and aims to launch the service by the end of the summer.