Hark has announced Handoff², the second version of the computer-use agent integrated into its Hark Pro application. According to the company, it completes web tasks approximately twice as fast as the original version, on average. These gains are attributed to reinforcement learning (RL) training, which penalized unnecessary turns and latency.
Handoff² works like a person in front of a screen: it clicks, types text, and scrolls pages live, instead of calling application programming interfaces (APIs). Hark introduced Handoff in August, before integrating it into Hark Pro in October. Since then, the company reports having recorded over 5.6 million browsing sessions and processed more than $20 million in purchases. Users entrust it with errands, flight tickets, dentist appointments, refund requests, warranty forms, or job applications.
How the Agent Gains Speed
Hark evolved the agent’s execution framework (its ‘harness’, meaning the environment in which the model acts) and the RL training method concurrently. The new framework allows the model to emit more actions per turn. The training then penalized superfluous turns, unnecessary reasoning tokens (the text fragments the model produces to think), and end-to-end latency.
The model thus learned three habits: grouping multiple actions, for example filling several form fields in a single turn; bundling repeated action sequences into reusable composite actions; and then reusing these composites throughout the task. Each request runs on a dedicated virtual computer, with its own browser, file system, and terminal. As savings accumulate each turn, gains increase on long, navigation-heavy tasks.
Figures Announced by the Publisher
Hark indicates an average speedup of approximately 2x compared to v1, up to 3.8 times faster for flight bookings and over 5 times for long searches. In one cited example, a 10-minute task drops to 3 minutes, and a 45-minute task to about 10 minutes.
On its internal navigation benchmark, which includes 478 tasks across 238 websites, Handoff² succeeds at 82%, compared to 73% for v1, with 46% less time per task. Hark ran Claude Opus 5.5 (max) in the same framework: according to these measurements, Handoff² remains at most 5 points behind this model. V1 was evaluated in its original framework. On its alignment benchmark, which checks, among other things, that the agent never confirms a purchase without approval and only asks for clarification when the request is genuinely ambiguous, the score rises from 90.3% to 93.8%. All these results are, according to the text, ‘self-reported’: no independent scores exist yet.
On the public Online-Mind2Web ranking, the original Handoff version scores 97.67%, and Yutori’s Navigator n1.5 scores 97.3%. Hark has not yet published any Handoff² results on this benchmark.
Access and Pricing
Handoff² is available in Hark Pro on the web, on iOS, and on Android. It operates solely in Hark’s cloud, with no local option or quantized version, and the company has announced neither a public API nor publication of the model weights. According to the pricing page, Hark Pro is free, the $20 per month plan offers twice the usage, and the $100 per month plan offers ten times the usage. All features are available across all plans.
