Hark previews its browser use agent for completing tasks
TechCrunch Ivan Mehta
Hark, fresh off a $700M round, showed off an AI agent that can click and type its way through websites like Target or LinkedIn on its own. It's a crowded field, but Hark claims its model predicts actions instead of just text, and undercuts rivals on price.
Brett Adcock's Hark just pulled back the curtain on Handoff, an AI agent built to do the grunt work of clicking through websites that never bothered to build an API. Think Target, Walmart, OpenTable, LinkedIn — the sites where you'd normally have to manually poke through menus to book a table or return a package. Handoff reads both the page's structure and what it visually looks like, then decides where to click or type, no different in spirit from the browser agents Google, OpenAI, and Anthropic have already shipped, or the ones coming out of funded startups like Browser Use and Aside.
What separates Hark's pitch, at least on paper, is the underlying model. Instead of predicting the next word the way a typical LLM does, Hark says its system predicts the next action — a click, a keystroke, placed at a specific spot on a page. It's a post-trained model for now, with a proper pre-training run planned later this year once the company has had time to sharpen its data pipeline and training setup. Whether that architecture actually holds up against a full context of messy real-world sites remains to be seen; right now it's a claim, not a benchmark anyone outside Hark has verified.
The demo Adcock posted has the agent assembling a flower bouquet from a loose, conversational request — something like leaving part of the order up to "the florist's choice." It's a decent showcase of handling vague human language, but the clip only shows a slice of the process, so there's no way to judge how it handles the harder, longer-tail failures that trip up basically every browser agent released to date: CAPTCHAs, login walls, sites that change their layout overnight.
Hark is also leaning hard on cost and speed as its wedge. The company says Handoff runs faster and considerably cheaper than heavyweight models like GPT-5.5 or Opus 4.8, which matters a lot if this thing is meant to run constantly in the background completing errands rather than being invoked occasionally for a single task. There's a waitlist open now, with a public release targeted for the end of summer — plenty of time for competitors to catch up, or for Hark to prove the action-prediction approach is more than just marketing language.
My take
Every well-funded agent startup claims it's faster and cheaper than the last one, and none of that means much until real users hit real paywalls, CAPTCHAs, and janky checkout flows at scale. A ten-second flower demo doesn't tell anyone whether this thing can survive a Ticketmaster queue, and until Hark shows the messy middle of a task, the $700 million valuation is doing more talking than the product is.”}}
Read more about this at: TechCrunch
