Hark, the artificial intelligence startup, launches Handoff, a browser-based agent designed to carry out online tasks by mimicking human behaviour, challenging conventional API-driven automation and highlighting new safety considerations.
Hark, the artificial intelligence startup founded by Brett Adcock, has introduced Handoff, a browser-based agent built to carry out online tasks by behaving much like a person at a computer. The system is aimed at work such as booking travel, ordering meals, shopping, making restaurant reservations, gathering information and handling recruiting tasks, all through direct interaction with websites rather than through application programming interfaces, or APIs.
The company argues that this approach is necessary because public APIs remain scarce across the web. Hark says fewer than 1 in 1,000 websites offer them, while its own research found that 74.9% of nearly 3 million minutes of observed screen time was spent inside browsers. That, the company says, makes browser navigation the central problem for AI assistants that need to do real work online.
To complete a request, Handoff sets up a dedicated virtual machine with its own browser, file system and terminal. Users can connect existing accounts so the agent can draw on stored addresses, preferences and history. Hark says the system is designed to cope with shifting layouts, pop-ups, adverts and other features that often trip up automation tools built around rigid scripts. The company says the model was post-trained with supervised fine-tuning and reinforcement learning, and is moving through later training stages ahead of a planned pretraining phase.
Hark is also making a performance case. It says Handoff posted a score of 97.7 on the Online-Mind2Web human evaluation leaderboard, ahead of GPT-5.4 and Claude Opus 4.8, and claims the system outperformed those rivals across several browser-use benchmarks. But VentureBeat noted that some of the comparisons were made against earlier model generations and that parts of Hark’s testing used its own harness, which makes independent verification more difficult. The strongest externally grounded claim remains the Online-Mind2Web result, though benchmark wins do not always translate into better performance in messy real-world use.
Cost is another part of the pitch. Hark says Handoff is cheaper to run than frontier models, listing token prices of $0.18 per million input tokens and $2.37 per million output tokens. The company believes that could attract businesses looking to automate repetitive digital work without paying top-tier AI prices, although token costs do not capture retries, infrastructure or failure handling.
Still, the biggest obstacle may be trust. A browser agent that merely summarises information poses limited risk; one that signs in, buys products or sends messages requires much stronger safeguards. Hark says Handoff can securely access connected accounts, but it has not yet publicly detailed those protections. The product remains in research preview, with wider availability planned by the end of summer.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
