Hark Enters Browser AI Race With New Handoff Agent

AI startup Hark has introduced Handoff, a computer-use agent designed to navigate websites and complete multi-step tasks through a browser much like a human user.

The research preview is the first detailed look at Hark’s product since the company raised a massive $700 million Series A in May 2026. That funding round valued the startup at $6 billion post-money, making Hark one of the most heavily funded new entrants in the consumer AI market.

Founded by Brett Adcock, who previously launched robotics company Figure AI and electric aircraft maker Archer, Hark is developing what it describes as a personalised intelligence platform combining proprietary AI models, software and eventually dedicated hardware.

Handoff Operates Websites Without Dedicated APIs

Handoff is built to perform long-running tasks directly through websites rather than relying on official application programming interfaces.

For each request, Hark creates a dedicated virtual computer containing a browser, file system and terminal. The agent then reads the page layout, interprets visual information and generates cursor and keyboard actions such as clicking, scrolling and typing.

Hark argues that browser control is essential because fewer than one in every 1,000 websites provides a public API. Even when an API is available, it may expose only a small part of what users can do through the normal website interface.

The company says Handoff can operate across services including DoorDash, Target, Walmart, Costco, OpenTable, LinkedIn, Booking.com and major airline websites.

Its demonstrated use cases include ordering food, comparing product prices, adding items to shopping carts, booking restaurant tables, researching information across multiple sources, finding job candidates and reserving flights or hotels.

In one demonstration, Hark CEO Brett Adcock asked the agent to assemble a flower bouquet based on several preferences, including a less precise instruction to add “some of the florist’s choice.” However, independent coverage noted that the video showed only part of the process, making it difficult to independently assess the reliability of the completed transaction.

Hark Claims Strong Browser-Benchmark Performance

Hark says Handoff achieved frontier-level results across three browser-use evaluations.

According to the company, the model reached the top position in the human-evaluated Online-Mind2Web leaderboard. Hark also claims that Handoff outperformed GPT 5.4 by an average of eight percentage points and Anthropic’s Opus 4.8 by two points across its evaluated benchmarks.

Some comparisons were conducted using Hark’s own testing framework, meaning the broader research community will need to independently reproduce the results before firm conclusions can be drawn.

The company also claims Handoff delivers comparable or better performance at roughly an order of magnitude lower token cost than competing frontier models. Its latency measurements exclude around 10 seconds of additional browser-processing time, an important consideration for tasks involving multiple web interactions.

Reinforcement Learning Helps the Agent Recover From Errors

Unlike conventional language models that primarily predict the next text token, Handoff is trained to generate the next computer action, such as moving a cursor to specific screen coordinates or entering text into a field.

Hark began development with supervised fine-tuning and then added reinforcement learning, allowing the agent to learn through successful and failed attempts on real browser tasks.

The company says reinforcement learning is particularly important because websites frequently change and can contain pop-ups, cookie banners, advertisements, bot detection systems and inconsistent navigation structures. These conditions make fixed browser scripts unreliable for general-purpose automation.

Hark is currently using a post-trained model but says it has moved into mid-training and plans to begin pre-training its own computer-use model later in 2026.

A Crowded Race for Control of the Browser

Handoff enters a rapidly expanding market that already includes browser and computer-use technology from OpenAI, Anthropic and Google, alongside startups such as Browser Use, Polar, Strawberry and Aside.

Hark’s large funding round gives it significant resources to compete. The company reportedly employed around 70 people at the time of its May financing and operated a data centre equipped with Nvidia B200 GPUs.

The next challenge will be proving that Handoff can complete sensitive, real-world transactions consistently without making incorrect purchases, exposing account information or taking unintended actions.

Hark has opened a waitlist for beta access and says its broader platform is scheduled to become available by the end of summer 2026. Until users can test the agent independently, its performance and cost advantages remain company claims rather than fully verified market results.