Robots Atlas>ROBOTS ATLAS
Artificial Intelligence

Hark unveils Handoff, a browser agent for real-world tasks

Sir Robot8 August 2026 · 4 min read
Hark unveils Handoff, a browser agent for real-world tasks

Hark on August 5, 2026 unveiled Handoff, an AI agent that operates websites on its own by spinning up a dedicated virtual computer — browser, file system and terminal — for every task. The product is available only in a closed preview for now, with a public launch planned for the end of summer 2026. The company says the model leads three browser-agent benchmarks while costing a fraction of rival systems.

Key takeaways

  • Handoff is a Computer Use agent that drives a live browser with cursor and keyboard actions
  • Every task gets its own virtual computer: browser, file system and terminal
  • Scores 97.7 on the Online-Mind2Web benchmark, ahead of GPT-5.5 (92.8) and Opus 4.8 (84.1)
  • Priced at $0.18 per million input tokens and $2.37 per million output — below one-tenth of GPT-5.5 rates
  • Public launch planned for late summer 2026, sign-up via hark.com

An agent that clicks instead of calling APIs

Handoff does not rely on the official programming interfaces of the sites it uses. Instead it reads the page structure and what is on screen, decides where to click or what to type, and returns cursor coordinates and keyboard signals that steer a live browser. Its makers stress that the model predicts the next action rather than the next text token — a different approach from classic chatbots.

For each request Hark spins up a dedicated virtual computer with a browser, a file system and a terminal. Inside it the user can safely log into their own accounts, and the agent preserves their preferences and history. The company showed Handoff ordering food on DoorDash and Uber Eats, comparing and booking flights on United, Delta and American, shopping at retailers such as Walmart, and recruiting and messaging candidates on LinkedIn.

Scores and cost as the core pitch

Hark builds its pitch on hard numbers. On Online-Mind2Web, which measures how well agents perform on real websites, Handoff scored 97.7 points — ahead of GPT-5.5 (OpenAI, 92.8), Claude Opus 4.8 (Anthropic, 84.1) and Gemini 2.5 Pro (Google, 69). The company claims it leads three browser-agent benchmarks at once.

AgentScore (pts)
Handoff (Hark)97.7
GPT-5.5 (OpenAI)92.8
Claude Opus 4.8 (Anthropic)84.1
Gemini 2.5 Pro (Google)69

Price is stressed just as hard. Handoff is set to cost $0.18 per million input tokens and $2.37 per million output, below one-tenth of GPT-5.5 rates of $5 and $30. That matters where an agent takes many steps on a single site and cost piles up with every action.

The world is full of AI assistants, but you'd never hire an assistant who couldn't use a computer.

Brett Adcock, founder of Hark.

A bet on the browser and an early-stage model

Hark justifies the browser as its battleground with user behavior — the firm says 74.9 percent of computer use happens in the browser. Handoff currently runs only on a post-trained model that combines supervised fine-tuning (SFT) with reinforcement learning aimed at recovering from failures on live sites. Full pre-training is scheduled for later in 2026, meaning the current results come from an incomplete version of the model.

74.9%of computer use happens in the browser — Hark's rationale for its battleground

The company positions Handoff directly against rival browser agents. Hark — which in May 2026 raised a $700 million Series A at a $6 billion valuation — is also building its own multimodal systems and a first hardware device.

Why it matters

Computer-use agents are one of the most contested directions in AI — the promise is large, but it has kept breaking on unreliability on unfamiliar sites and the high cost of repeated steps. Handoff attacks both at once, claiming the top score on real websites and a price around one-tenth of rivals. If those figures hold beyond the benchmark, they shift the debate from whether an agent will click the right button to whom we will trust with access to our accounts and payments.

What's next

  • The public platform launch is planned for the end of summer 2026 — a waitlist is open now at hark.com
  • Full pre-training of the model is due later in 2026 and will show whether the preview lead holds in a production release
  • Hark is developing a first hardware device in parallel, pointing to a plan to pair the agent with its own hardware

Sources

Share this article