Introducing Hark Handoff
hark.com ● Covered by 2 sources
Hark just launched Handoff, an AI agent that browses the web and finishes tasks like ordering food or booking flights, all on its own. It claims top marks against GPT and Opus rivals, and it's built because most sites have no API to plug into.
Based on reporting by hark.com — 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's big pitch is simple: most of what people do on computers happens inside a browser, and almost none of that is reachable through clean software hooks. The company says its own usage survey, covering close to three million screen minutes, found nearly 75 percent of that time sitting in a browser tab. People are online more than six hours a day, generating trillions of site visits annually across roughly 300 million domains, and Hark points out that fewer than one in a thousand of those sites expose a public API. Places like DoorDash, Target, Walmart, OpenTable, Zillow, and LinkedIn simply don't offer one. So if you want an agent that can actually get things done for someone, it has to click, type, and scroll through the same messy interfaces a human does.
That's the job Handoff is built for. Every request spins up its own virtual computer, complete with a browser, file system, and terminal, rather than running a script against a fixed list of sites. It can log into a person's existing accounts and act with their saved addresses, payment info, and preferences already loaded in. Hark shows it ordering food on DoorDash, comparison shopping and checking out on Target, booking a table on OpenTable, messaging job candidates on LinkedIn, cross-referencing Reddit and review sites for research, and reserving flights or hotels on sites like United and Booking.com.
On benchmarks, Hark says Handoff takes the top spot on the Online-Mind2Web human evaluation leaderboard, beating GPT 5.4 by an average of 8 points and Opus 4.8 by 2 points, with a wider gap over Gemini-3.5 Flash and Gemini-2.5 Pro. It also reports strong results on WebTailBench and its own internal evaluations, all run through a harness Hark built in-house. The company frames pricing as another edge, saying Handoff runs at an order of magnitude lower cost per token than the frontier models it's compared against, though browser latency still adds roughly ten seconds on top of the model's own per-turn response time.
The technical story behind Handoff is a six-month climb through training stages. Hark started with post-training because reliable data for internet-based computer use is scarce, and post-training let the team iterate fast on pipelines and evals before committing to anything heavier. It has since moved into mid-training, with pre-training planned later this year. Supervised fine-tuning and reinforcement learning both contributed to the gains, according to Hark's own pass@1 charts, with RL doing more of the work on tasks the base model already handled well and SFT carrying more weight on harder ones.
Hark frames all of this as a step toward computers that use computers on a person's behalf, rather than just holding their attention. The full platform is slated for release by the end of summer, and Hark is currently taking sign-ups for early access.
My take — AI-written commentary, not fact-checked reporting
The interesting number here isn't the benchmark score, it's that fewer than one in a thousand sites bother with a public API — which means any agent company betting on clean integrations is building on sand, while whoever gets browser automation right owns the real surface area of the internet. Beating GPT and Opus on a leaderboard is nice marketing, but the actual test is whether Handoff survives six months of DoorDash redesigning its checkout flow without falling over. Cheap and fast matters more than clever here, because nobody wants an assistant that takes longer than doing the task themselves.
Read more about this at: hark.com
