Hark, the AI startup founded by Figure AI founder Brett Adcock, has unveiled Handoff, a computer-use agent designed to complete online tasks by operating websites like a human user.
The company says Handoff can handle tasks such as ordering food, booking flights, shopping online, making restaurant reservations, researching information and managing recruiting workflows. Instead of relying on website APIs, the agent interacts directly with browsers by clicking, typing and scrolling.
Hark argues this approach is necessary because most websites do not offer public APIs that allow AI systems to perform tasks. The company estimates that fewer than one in 1,000 websites provide public APIs, leaving AI agents with limited ways to interact with many everyday services.
Hark said it surveyed users across nearly 3 million minutes of screen time and found that 74.9% of that time was spent inside browsers, making web navigation a key challenge for AI assistants
How Handoff works
For each request, Hark creates a dedicated virtual computer environment with its own browser, file system and terminal. Users can connect their existing accounts, allowing Handoff to use saved information such as addresses, preferences and history.
Unlike traditional automation tools built around fixed workflows, Hark says Handoff is designed to adapt to changing websites, pop-ups, advertisements and other obstacles that make the internet difficult for automated systems.
Hark said the agent was post-trained through supervised fine-tuning and reinforcement learning. The company has since moved into mid-training and plans to begin pretraining later this year. Handoff’s output consists of actions such as cursor movements and keyboard inputs rather than text responses alone.
Benchmark claims and performance questions
Hark claims Handoff achieved leading results on several browser-use benchmarks, including the Online-Mind2Web human evaluation leaderboard.
The company said Handoff scored 97.7 on Online-Mind2Web, ahead of GPT-5.4’s 92.8 and Claude Opus 4.8’s 84.1. Hark also reported that its system outperformed GPT-5.4 by 8 points on average and Opus 4.8 by 2 points across three browser benchmarks.
However, some comparisons come with limitations. VentureBeat noted that Hark’s testing compared Handoff against earlier-generation models. Hark also said its WebTailBench and internal evaluations were conducted using the company’s own testing harness.
The company’s most externally grounded result comes from Online-Mind2Web, where Hark says Handoff currently holds the top score. But benchmark leadership does not necessarily guarantee better performance across every real-world situation, and the broader results have not been independently reproduced.
Lower costs could be Hark’s biggest advantage
Beyond performance, Hark is positioning Handoff as a cheaper alternative to major AI models.
The company says its agent delivers performance at a significantly lower token cost than competing frontier models. Hark listed rates of $0.18 per million input tokens and $2.37 per million output tokens, compared with higher rates for models from major AI providers.
If the pricing and reliability claims hold up, Handoff could appeal to businesses looking to automate repetitive digital work without paying premium AI infrastructure costs. However, token rates alone do not account for the total cost of completing a browser task, including retries and supporting infrastructure.
The challenge: earning user trust
Browser agents face a different hurdle from chatbots: users must trust them with real actions.
An AI assistant that summarizes information is relatively low risk. An agent that logs into accounts, selects purchases or sends messages needs stronger safeguards around security, privacy and mistakes.
Hark says Handoff can securely access connected accounts, but the company has not publicly detailed the safeguards protecting those accounts. Handoff also remains a research preview, with broader availability planned by the end of summer.
The next phase will show whether users are comfortable handing over everyday tasks to autonomous agents. For Hark, the technology may be impressive, but widespread adoption will depend less on benchmark scores and more on whether people trust an AI system to make decisions inside their personal accounts.
Related reading: Hark is not alone in pushing AI beyond chat and into autonomous work. See how OpenAI’s ChatGPT Work handles multi-step workplace tasks using connected business tools and data.