AI startup Hark has released Handoff², the second generation of its web-navigating computer-use agent available inside the Hark Pro application. The updated agent executes web-based tasks approximately two times faster than its predecessor by directly interacting with browser elements, screen coordinates, and virtual operating systems.
To achieve the performance boost, Hark co-designed its execution harness and reinforcement learning (RL) training recipe. The model was trained to output multiple actions per turn while explicitly penalizing latency, excess reasoning tokens, and unnecessary steps. According to the company, the Hark Pro service has handled over 5.6 million consumer browsing sessions and processed $20 million in consumer transactions since launching its initial agent preview.
Handoff² is deployed natively within Hark Pro’s web and mobile applications under tiered subscription pricing. Hark reported that internal benchmark comparisons place Handoff² within five performance points of Claude Opus 5.5 running in the same custom harness.
Why it matters
Agent developers can achieve massive speed and cost efficiency by using reinforcement learning to penalize idle turns and excess thinking tokens.
Computer-use models operating directly on UI coordinates are proving commercially viable for consumer-facing transaction automation.
Source: marktechpost.com



