TLDRocket
Sign in

Hark previews its browser use agent for completing tasks

TechCrunch Ivan Mehta ● Covered by 2 sources

Hark just showed off Handoff, an AI agent that browses sites like Target or LinkedIn to do tasks for you. It claims to predict actions instead of just text, and says it's cheaper and faster than rivals.

Based on reporting by TechCrunch, Ivan Mehta — read the original for the full story.

Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error

Hark, the startup that pulled in $700 million in a Series A back in May, is now showing off its first product: an agent called Handoff that's built to operate a web browser the way a person would. Point it at Target, Walmart, OpenTable, or LinkedIn — none of which offer the kind of official API that makes automation easy — and the agent is supposed to figure out the page layout on its own, deciding where to click and what to type based on both the site's structure and what it visually looks like.

The pitch itself is nothing new. Browser agents that take a plain-language command and go complete a task, whether that's grabbing takeout, booking a flight, filing a return, or hunting down information across multiple sites, have been showing up from a string of companies for a while now. What Hark is leaning on to stand out is a claim about how the model itself works. CEO Brett Adcock demoed the assistant assembling a flower bouquet from a loose request, including handling a vague instruction like letting the florist choose the arrangement. The catch is the demo only showed a slice of the process, so there's no real way to judge yet how well it actually performs end to end.

Under the hood, Hark says Handoff currently runs on a post-trained model, with plans to move to pre-training later this year — a sequencing choice the company says lets it iterate faster on its data pipeline and training setup before locking things in. The more interesting technical claim is that instead of predicting the next word the way a typical large language model does, Handoff is built to predict the next action: a click, a keystroke, placed at a specific spot on the screen.

That's a crowded field to be making bold claims in. Google, OpenAI, and Anthropic are all pushing their own computer-use systems, and a batch of VC-backed startups — Browser Use, Polar, Strawberry, and Aside among them — are chasing the same idea of browser-driven task automation. Hark's counter is speed and price, arguing Handoff beats the competition on both and undercuts the cost of running models like GPT 5.5 or Opus 4.8. For now, that's a claim rather than a demonstrated result. The company has opened a waitlist and is targeting a release by the end of summer.

My take — AI-written commentary, not fact-checked reporting

Calling your model's approach fundamentally different because it predicts actions instead of tokens sounds impressive right up until someone asks for a benchmark instead of a bouquet video. A funding round that size buys attention, but it doesn't buy proof, and in a field already stuffed with Google, OpenAI, Anthropic, and a pack of scrappier startups all claiming the same speed-and-cost edge, the company that actually publishes a full, unedited task run first is the one worth believing.

Read more about this at: TechCrunch

Related stories

The daily briefing

Every AI story that matters, in your inbox by 8am.

TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.