📰 Key Takeaways

Hark, the startup that raised $700M in a Series A this past May, launched its browser-use agent Hark Handoff today, positioning it as an efficient way to complete all kinds of tasks in a browser. The company says Handoff can operate smoothly on websites without official APIs, including Target, Walmart, OpenTable, and LinkedIn, by having the agent parse both the website’s structure and its visual layout at the same time, so it can figure out which button to click or where to type. The concept is similar to the wave of browser agents that came before it: you give it an instruction, and the agent can order food or coffee for you, book tickets, handle returns, buy groceries, reserve a table, or gather information from multiple sources to do research on your behalf. In the demo video, CEO Brett Adcock showed the agent being told to put together a bouquet with flowers the user specified, and it was able to understand vague instructions like “let the florist pick some of the flowers” — though the video only shows part of the process, so it’s hard to fully judge how well it actually works. The company says this release runs on a post-trained model, and it plans to move to pretraining by the end of the year, which should speed up how fast it can improve its data pipeline, training infrastructure, and related tech. Hark also claims that unlike traditional LLMs that predict the next token, its model directly predicts the “next action” — like clicking somewhere specific on the screen or typing on the keyboard. Other players currently working on computer-use agents include Google, OpenAI, Anthropic, and startups like Browser Use, Polar, Strawberry, and Aside. Hark says Handoff is faster than the competition and costs far less than models like GPT 5.5 and Opus 4.8. It’s on a waitlist now, with a full launch planned before summer’s out.

💬 JudyAI Lab Take

We often assume that automating anything requires an official API first, but Hark flips that: by having the agent read both the page’s DOM structure and its visual layout at the same time, it can operate directly on sites like Target, Walmart, and OpenTable that don’t expose APIs at all — basically routing around the “does it have an API” gate entirely.

This points to a division of labor that’s becoming clearer across the industry: language models handle understanding and decision-making, while the actual operating layer gets handed off to specialized agents that can read both DOM and screen visuals, each trained and optimized separately. Hark’s claim that it predicts the “next action” instead of just chaining tokens — redesigning the model’s objective specifically for operating tasks — could make browser agents more reliably effective than just bolting a general-purpose LLM onto the problem, while also pushing costs down. It’s a concrete example of splitting post-training and pretraining investment into stages.

Worth keeping an eye on how these agents actually hold up in terms of error tolerance and reproducibility on real websites, rather than just judging from demo clips.

📅 Source Info

🔗 Further Reading