Skip to main content
Hark Pro's new 2x faster Handoff² feature, with original Handoff scoring 97.67%, improves workflow efficiency.

Editorial illustration for Hark Releases 2x Faster Handoff² for Hark Pro; Original Handoff Scores 97.67%

Hark Handoff² Doubles Speed for Web Task Agent

Hark Releases 2x Faster Handoff² for Hark Pro; Original Handoff Scores 97.67%

• 4 min read

Hark shipped Handoff² this week, the second version of the computer-use agent built into its Hark Pro app, and the company says it completes web tasks roughly twice as fast as the original. Handoff first showed up as a preview in August, went live in Hark Pro in October, and has since run more than 5.6 million browsing sessions and processed over $20 million in purchases, according to Hark. The agent doesn't call APIs. It clicks, types, and scrolls on live websites the same way a person would, which is why speed and wasted motion matter so much to how it's graded.

Hark credits the jump to reinforcement learning that specifically punished extra turns and latency, not just wrong answers. That's a different training target than most agent labs chase, where accuracy usually gets all the attention and speed is an afterthought. On Hark's own 478-task internal benchmark, Handoff² hits 82% completion while running about twice as fast end to end as v1.

The company also reports a jump on its internal alignment benchmark, from 90.3% to 93.8%. All of these numbers come from Hark itself, with no outside lab having verified them yet.

Hark has released Handoff², the second version of the computer-use agent inside its Hark Pro app. The Hark Handoff 2 computer-use agent finishes web tasks about 2x faster than the original, on average. Hark states that the gains come from reinforcement learning (RL) that penalized wasted turns and latency.

Why this matters

Speed claims without a matching benchmark number are easy to make and hard to verify. Hark says Handoff² runs tasks twice as fast as the original, but the original's 97.67% on Online-Mind2Web is the only hard figure on the table, and that score belongs to the slower model. Until Hark posts a Handoff² result on the same leaderboard, there's no way to know whether that 2x speed gain came with a 2x accuracy trade-off or none at all.

For developers and founders building browser agents, the RL approach itself is worth noting: penalizing wasted turns and latency directly, rather than just rewarding task completion, is a sensible training lever if you're trying to ship something usable in production. But the lack of a public API and the absence of a fresh eval score means outside teams can't test the claim themselves right now. Worth watching whether Hark publishes a Handoff² leaderboard entry, and whether an API shows up for teams that want to benchmark it against Claude, GPT, or other computer-use agents directly.

Common Questions Answered

How much faster is Handoff² compared to the original Handoff agent?

Handoff² completes web tasks roughly twice as fast as the original Handoff agent on average. According to Hark, these speed improvements come from reinforcement learning that specifically penalized wasted turns and latency in the system.

What is the accuracy score of the original Handoff on Online-Mind2Web?

The original Handoff achieved a 97.67% accuracy score on the Online-Mind2Web benchmark. This is currently the only hard benchmark figure publicly available for Hark's computer-use agents.

How does Handoff interact with websites without using APIs?

Handoff clicks, types, and scrolls on live websites the same way a person would, rather than calling APIs directly. This approach allows it to work with any website without requiring special integrations or API access.

What metrics demonstrate Handoff's real-world usage since its October launch?

Since going live in Hark Pro in October, Handoff has run more than 5.6 million browsing sessions and processed over $20 million in purchases. These figures showcase the agent's adoption and practical application in handling web-based transactions.

What verification concern exists regarding Handoff²'s claimed 2x speed improvement?

Without Handoff² posting results on the same Online-Mind2Web leaderboard as the original model, there is no way to verify whether the 2x speed gain came with a corresponding accuracy trade-off or maintained the same performance level. Speed claims without matching benchmark numbers are difficult to independently verify.

LIVE12:33Hark Releases 2x Faster Handoff² for Hark Pro; Original Handoff Scores 97.67%