Hark previews its browser use agent for completing tasks

Startup Hark, which successfully secured $700 million in Series A funding back in May, has officially launched its new agent named Hark Handoff, engineered to operate web browsers efficiently and execute tasks on behalf of users.

The company asserts that the Handoff agent can seamlessly navigate through websites lacking official APIs, mentioning platforms such as Target, Walmart, OpenTable, and LinkedIn. According to the firm, the agent inspects website structures and visual data to determine whether it needs to trigger a button click or input text data.

The core concept mirrors various browser-based agents introduced previously. Users can issue a command, and the software handles execution, ranging from ordering food or coffee, booking travel tickets, processing returns, shopping for essentials, reserving restaurant tables, to performing research across multiple sources.

<

Regarding the underlying architecture, the company stated that this initial release relies on a post-trained model, with plans for pre-training later in the year. Hark claims this strategy allows them to refine their data pipeline, training infrastructure, and techniques at a faster pace. Furthermore, Hark highlights that unlike traditional large language models (LLMs) that predict the next token, their model predicts the next action, such as a specific mouse click or keyboard input at a precise location.