Hark, the startup that vacuumed up a $700 million Series A in May, finally showed its hand. The company launched Hark Handoff, a browser-based agent designed to complete real-world tasks on sites that lack official APIs — Target, Walmart, OpenTable, LinkedIn. The pitch is straightforward: you issue a command, and Handoff parses a website’s visual structure and layout to figure out where to click and what to type, effectively mimicking human interaction.
CEO Brett Adcock shared a video demo where Handoff builds a custom flower bouquet, even interpreting fuzzy instructions like “some of the florist’s choice.” But here’s the catch — the demo is heavily edited, showing only a sanitized slice of the process. We don’t see the failures, the misclicks, or the moments the agent stalls on a dynamic pop-up. That’s the real test for any computer-use agent, and Hark is not yet showing its work.
The technical approach is what sets Hark apart, at least on paper. Rather than relying on a large language model predicting the next token, Hark claims its model predicts the next action — a click or a keystroke at a specific coordinate. The company is launching with a post-trained model and plans to move to pre-training later this year, a sequencing maneuver they say lets them refine their data pipeline faster. It’s an ambitious architectural bet in a field that’s already crowded with players like Google, OpenAI, and Anthropic, not to mention VC-backed upstarts like Browser Use and Polar.
Hark is making aggressive claims about speed and cost, saying Handoff runs circles around competitors and undercuts the pricing of frontier models like GPT 5.5 and Opus 4.8. But without independent benchmarks or unedited demos, those claims are just marketing. The waitlist is open, with a full release promised by summer’s end. The $700 million question is whether Handoff can handle the messy, unpredictable web outside of a controlled demo — because that’s where browser agents have consistently stumbled.
