Wink Pings

Hark's Computer Use Agent Is 2x Faster, With a Standout 82% Task Completion Rate

Hark has released Handoff², which delivers twice the average speed of its v1 predecessor and hits an 82% completion rate across 478 internal web tasks. It learns to execute actions in batches through reinforcement learning with latency penalties, though the limitations of this internal benchmark need to be properly acknowledged.

Hark has launched Handoff², the computer-use agent built into Hark Pro. As reported by MarkTechPost author Michal Sutter, it runs at twice the average speed of the previous iteration, and achieves an 82% task completion rate across a set of 478 in-house web tasks.

What computer-use agents do is very straightforward: view web pages, click buttons, fill out forms, and walk through workflows across different pages. It does not call website APIs, and completes tasks fully via interface operations. This category of products has two worst-case pitfalls: slowness, and wrong clicks. Slowness can drag a chain of tiny steps out to several minutes, while misclicks can send the entire workflow straight to a dead end.

The major upgrade of Handoff² centers on batched actions. Per the report, it uses reinforcement learning with latency penalties to learn to combine multiple actions for concurrent execution. The key here is the latency penalty: every step of waiting is counted as a cost, so the model is strongly incentivized to cut down unnecessary back-and-forth trips. For web operations, this design aligns far better with real user experience than simply raising single-step accuracy, since users care about how long the entire task takes to finish from their perspective.

But that 82% figure needs to be viewed in its full context. The 478 test tasks are all internal web tasks, part of a private internal benchmark, not a public open benchmark. Internal task sets are usually more tailored to the exact scenarios the product is built to sell, and it is far easier to curate demo-ready samples from these sets. The metric proves Handoff² is far more capable than the v1 version, but it does not directly mean the agent can handle arbitrary workflows across the open public internet.

For existing Hark Pro users, this update solves a core efficiency pain point: the exact same web workflow now takes half the average waiting time to finish. The real performance to watch for is its stability when dealing with complex pages, persistent login states, unexpected pop-ups, and unplanned error branches. Doubled speed is a great milestone, but whether the stated completion rate can hold up in messier, uncurated real-world environments is the next critical threshold for computer-use agents to cross.

Original link: Hark Releases Handoff², a 2x Faster Computer-Use Agent for Hark Pro

发布时间: 2026-11-16 17:08