AINEWS
Handoff, a browser agent that drives sites like a person, ships behind a waitlist by end of summer. The $700M round buys entry into a category where Google, OpenAI, and Anthropic already ship.
Hark, a startup that raised $700 million in May at a $6 billion valuation with no shipping product, this week previewed Handoff, a browser-use agent: software that drives a real web browser the way a person would, clicking "buy" on Target, booking a table on OpenTable, and completing tasks that today require a human. The browser-agent category it is entering already has Google, OpenAI, and Anthropic shipping, plus four named startups (Browser Use, Polar, Strawberry, and Aside).
Handoff was shown on Tuesday by CEO Brett Adcock in a video that builds a flower bouquet on a florist's site. TechCrunch noted the video shows only part of the bouquet-building process, so the claim that the agent finished the job is not visually verifiable from the public material. The preview is real. The proof is not.
Hark says Handoff does not use the dominant next-token-prediction paradigm. Instead, Adcock told TechCrunch the model is post-trained to predict "the next action," a click or a keyboard input at a specific place on a page. That framing is the company's argument about how browser-use work should be done. No independent benchmark, leaderboard, or third-party evaluator is named in the source material, and Adcock's tweet the same day that Handoff had been "independently verified as the best internet-use model ever built" did not name the verifier.
Hark says Handoff is faster than rivals and costs less than "GPT 5.5" and "Opus 4.8," the company's terms for the latest models from OpenAI and Anthropic. Those are first-party marketing claims. The category is not.
Hark's $700 million Series A, covered in May by TechCrunch, Bloomberg, and the company's own press release, was led at a $6 billion valuation before any product shipped. Forbes asked, bluntly, whether a $6 billion valuation with no product makes sense. The question is not whether Hark is worth $6 billion. The question is what kind of category requires a $700 million first round to enter in 2026.
The $700 million price breaks into three forces. Talent: browser-use agents process both site text and on-screen pixels, so they sit on top of large multimodal models. The people who can train them are the same people frontier labs are paying to keep, and a first round of this size buys runway through a multi-year training effort. Compute: Adcock says the model is post-trained now and that the company plans to pretrain its own model later in 2026. Pretraining a competitive multimodal model from scratch in 2026 is, by any published estimate, a nine-figure compute bill. Demo: showing a working product, even a partial one, is the only way to recruit engineers and to set the bar against incumbents who already own the rest of the stack.
A $700 million first round at a $6 billion valuation needs a public handle for the next round. A working demo, even one that does not finish a flower order, gives a Series B investor something to point at when the round is being priced.
Handoff ships behind a waitlist by the end of summer 2026, according to the company. The next time readers will see the $700 million justified on a number is when Hark publishes the benchmark sheet it has not yet released.