Hark Handoff scores 97.7 on OM2W with API-free web navigation

On August 5, 2026, AI startup Hark previewed Handoff, a browser-use agent that autonomously navigates web interfaces to execute tasks like booking travel or conducting candidate research. Founded by Brett Adcock, the company is shifting from traditional API-based automation to visual interpretation. For teams evaluating what AI agents are and how they work, Handoff offers a production alternative that interacts directly with unmodified user interfaces.

Action Prediction and Execution

Unlike standard models that predict the next text token, Handoff is a Computer Use Agent (CUA) trained to predict specific interface actions, such as screen coordinate clicks and keyboard inputs. The agent executes these actions inside a dedicated virtual computer spun up for each task. This isolated environment contains its own browser, filesystem, and terminal.

This architecture bypasses the need for underlying integrations. Hark internal research found that fewer than 1 in 1,000 websites maintain public APIs. By navigating the visual layer directly, Handoff allows users to connect existing accounts to complete transactions autonomously using saved addresses and payment methods.

Benchmark Results and Cost

Hark evaluates Handoff using the Online-Mind2Web (OM2W) benchmark, an industry standard for web agents. The current post-trained model achieved a score of 97.7.

Model	OM2W Score
Hark Handoff	97.7
GPT-5.4	92.8
Claude Opus 4.8	84.1
Gemini 2.5 Pro	69.0

When evaluating and testing AI agents, speed and token costs dictate operational feasibility. Hark reports a per-turn latency of 0.8 seconds. Input tokens are priced at $0.18 per million, while output tokens cost $2.37 per million. The company plans to release a pre-trained model later in 2026.

If you build task automation pipelines, Handoff presents a shift away from maintaining brittle API integrations toward visual interface control. You will need to weigh whether your workflows require the exact state guarantees of a structured API or if visual navigation can handle the edge cases of your target domains.