Hark, the AI startup based by Figure AI founder Brett Adcock, has unveiled Handoff, a computer-use agent designed to finish on-line duties by working web sites like a human person.
The firm says Handoff can deal with duties akin to ordering meals, reserving flights, purchasing on-line, making restaurant reservations, researching info and managing recruiting workflows. Instead of counting on web site APIs, the agent interacts instantly with browsers by clicking, typing and scrolling.
Hark argues this method is critical as a result of most web sites don’t provide public APIs that permit AI programs to carry out duties. The firm estimates that fewer than one in 1,000 web sites present public APIs, leaving AI brokers with restricted methods to work together with many on a regular basis companies.
Hark mentioned it surveyed customers throughout practically three million minutes of display time and located that 74.9% of that point was spent inside browsers, making internet navigation a key problem for AI assistants
How Handoff works
For every request, Hark creates a devoted digital pc atmosphere with its personal browser, file system and terminal. Users can join their current accounts, permitting Handoff to make use of saved info akin to addresses, preferences and historical past.
Unlike conventional automation instruments constructed round mounted workflows, Hark says Handoff is designed to adapt to altering web sites, pop-ups, commercials and different obstacles that make the web tough for automated programs.
Hark mentioned the agent was post-trained by supervised fine-tuning and reinforcement studying. The firm has since moved into mid-training and plans to start pretraining later this 12 months. Handoff’s output consists of actions akin to cursor actions and keyboard inputs slightly than textual content responses alone.
Benchmark claims and efficiency questions
Hark claims Handoff achieved main outcomes on a number of browser-use benchmarks, together with the Online-Mind2Web human analysis leaderboard.
The firm mentioned Handoff scored 97.7 on Online-Mind2Web, forward of GPT-5.4’s 92.Eight and Claude Opus 4.8’s 84.1. Hark additionally reported that its system outperformed GPT-5.Four by Eight factors on common and Opus 4.Eight by 2 factors throughout three browser benchmarks.
However, some comparisons include limitations. VentureBeat famous that Hark’s testing in contrast Handoff in opposition to earlier-generation fashions. Hark additionally mentioned its WebTailBench and inner evaluations have been performed utilizing the corporate’s personal testing harness.
The firm’s most externally grounded end result comes from Online-Mind2Web, the place Hark says Handoff presently holds the highest rating. But benchmark management doesn’t essentially assure higher efficiency throughout each real-world state of affairs, and the broader outcomes haven’t been independently reproduced.
Lower prices might be Hark’s largest benefit
Beyond efficiency, Hark is positioning Handoff as a less expensive various to main AI fashions.
The firm says its agent delivers efficiency at a considerably decrease token price than competing frontier fashions. Hark listed charges of $0.18 per million enter tokens and $2.37 per million output tokens, in contrast with increased charges for fashions from main AI suppliers.
If the pricing and reliability claims maintain up, Handoff may attraction to companies seeking to automate repetitive digital work with out paying premium AI infrastructure prices. However, token charges alone don’t account for the entire price of finishing a browser activity, together with retries and supporting infrastructure.
The problem: incomes person belief
Browser brokers face a distinct hurdle from chatbots: customers should belief them with actual actions.
An AI assistant that summarizes info is comparatively low danger. An agent that logs into accounts, selects purchases or sends messages wants stronger safeguards round safety, privateness and…
