Hark is moving into browser-based task automation with Hark Handoff, an agent built to use websites directly and carry out practical online jobs for users. The startup, which raised $700 million in Series A funding in May, says the system can work across familiar services even when those services do not provide official APIs.
What Hark Handoff Is Built To Do
Hark Handoff is presented as a browser use agent: a system that receives an instruction, opens or navigates websites, and takes actions through the browser interface. Instead of relying only on a formal software connection, Hark says the agent can interpret website structure and visual data to decide what to do next.
That means Handoff is meant to identify when it should click buttons, type information, or continue through a website flow. The company specifically points to sites including Target, Walmart, OpenTable, and LinkedIn as examples of places the agent can navigate without official APIs.
The broader promise is familiar in the current wave of browser agents. A user gives a command, and the agent attempts to complete a real-world online task. Hark describes use cases such as ordering food or coffee, booking travel tickets, filing returns, shopping for essentials, reserving a table at a restaurant, or researching across sources.
The Demo Shows The Ambition And The Limits
In a video demo, Hark CEO Brett Adcock showed the assistant handling a request to build a bouquet. The command included flowers selected by the user and also allowed a looser instruction, including the phrase “some of the florist’s choice.”
That example matters because many browser tasks are not just a sequence of fixed clicks. They often include vague user preferences, incomplete instructions, and choices that must be made inside a website’s existing interface. Hark is positioning Handoff as an agent that can translate that kind of request into browser actions.
At the same time, the demo leaves an important caveat. The video shows only part of the process, so it does not provide enough evidence to judge how consistently the assistant performs from start to finish. For browser automation, partial success is not the same as reliable completion, especially when tasks involve purchases, bookings, reservations, or account-based workflows.
How Hark Describes Its Model Strategy
For this release, Hark says it is using a post-trained model. The company also says it plans to pre-train later this year.
Hark’s explanation is that the current approach lets it improve its data pipeline, training infrastructure, and techniques more quickly. In plain terms, the company is presenting this preview as both a product step and a training step: it wants the agent in front of users while it continues refining how the system learns and acts.
The company also draws a distinction between its approach and large language models. Hark says that instead of predicting the next token, its model can predict the next action, which could involve a clock or a keyboard input at a specific place.
That framing is central to the browser agent category. A chat model may produce useful text, but a browser-use agent has to decide where to act inside a live interface. The key challenge is not only understanding a request, but converting that request into the right sequence of interactions on pages that may not be designed for automation.
A Crowded Race For Computer-Use Agents
Hark is not entering an empty field. The source names Google, OpenAI, and Anthropic among companies working on computer-use agents. It also lists VC-funded startups focused on browser-based task automation, including Browser Use, Polar, Strawberry, and Aside.
That competition shows why Hark is emphasizing speed, cost, and practical website coverage. The company says Handoff is faster than the competition and costs much less than other models like GPT 5.5 and Opus 4.8.
Those claims are significant, but the available information does not include independent benchmarks or detailed comparisons. For readers following AI agents, the practical question remains straightforward: can a browser agent complete common online tasks accurately, affordably, and with enough reliability to be useful beyond a controlled preview?
What Comes Next
Hark has opened a waitlist for Hark Handoff. The company plans to release the platform by the end of the summer.
Until users can test it directly, the clearest facts are the product direction and the claims Hark is making. Handoff is designed to act through the browser, work on sites without official APIs, and interpret both structure and visual cues. It is also arriving in a market where major AI companies and startups are all pursuing agents that can operate software on behalf of users.
If Hark can turn the preview into a reliable product, Handoff could become part of a broader shift from AI systems that answer questions to AI systems that complete tasks. For now, it is a notable entrant in the browser agents race, with a waitlist open and more proof still needed once the platform is released.