Hark previews its browser use agent for completing tasks
Following a massive $700 million Series A funding round this past May, startup Hark is stepping into the spotlight with the unveiling of Hark Handoff. This new browser-based agent is designed to autonomously navigate the web to execute complex user tasks, positioning itself as a high-efficiency alternative in a crowded field of automation tools.
Bridging the API Gap
Hark Handoff distinguishes itself by its ability to interact with websites that lack official APIs. By analyzing both the underlying structure of a webpage and its visual layout, the agent determines when to trigger clicks or input data. The company suggests this capability allows it to operate across diverse platforms, including:
- Retailers: Target and Walmart
- Services: OpenTable
- Professional Networks: LinkedIn
The agent is built to handle a wide array of requests, from booking travel and filing tax returns to shopping for household essentials or conducting multi-source research. In a recent demonstration, CEO Brett Adcock showcased the agent’s ability to handle nuanced instructions, such as building a custom floral bouquet that included both specific requests and subjective "florist’s choice" items.
"Unlike large language models (LLMs) predicting the next token, our model can predict the next action — which could be a click or a keyboard input at a specific place."
A New Approach to Automation
While the current version of Handoff utilizes a post-trained model, Hark intends to transition to a pre-trained model later this year. The company believes this strategy allows for more agile refinement of its data pipelines and training infrastructure.
Hark is entering a highly competitive landscape, facing off against tech giants like Google, OpenAI, and Anthropic, as well as specialized startups such as Browser Use, Polar, Strawberry, and Aside. Despite the competition, Hark is making bold claims regarding its performance, asserting that Handoff is not only faster than rival offerings but also significantly more cost-effective than advanced models like GPT 5.5 and Opus 4.8.
Looking Ahead
While the initial video demo provided a glimpse into the agent's capabilities, the full scope of its real-world effectiveness remains to be seen. Interested users can currently sign up for a waitlist on the company’s website, with a broader platform release slated for the end of the summer.