Hark’s New Browser Agent Chases Speed With Smarter Actions

Hark has released Handoff², the second version of the computer-use agent inside its Hark Pro app. Hark says the new agent completes web tasks about 2x faster than the original Handoff, with the biggest gains appearing on long, browsing-heavy jobs.
Handoff² operates websites the way a person does. It clicks, types and scrolls on live pages instead of calling APIs, which lets users send it tasks such as grocery orders, flight bookings, dentist appointments, refunds, warranty forms and job applications.
The headline numbers are strong, but they come with an important qualification: every Handoff² result is self-reported, and no third-party score exists yet.
Speed Comes From Fewer, Better Turns
Hark co-evolved two parts of the system for this release: the agent harness and the reinforcement learning training recipe. The new harness lets the model emit more actions per turn, while training penalized excess turns, unnecessary thinking tokens and end-to-end latency.
That training produced three habits. Handoff² can batch actions, group repeated action patterns into reusable composite actions, and reuse those composites as the task continues. Instead of treating every click or keystroke as a separate decision, the agent can carry out more of the work in each turn.
Those savings add up across a task. Hark reports that Handoff² is about 2x faster end to end than v1 on average, with flight booking up to 3.8x faster and long research tasks more than 5x faster. A 10-minute task can drop to 3 minutes, while a 45-minute task can take about 10 minutes.
For teams building browser agents, the release offers a clear case study in training for speed, not only accuracy. The model still needs to complete the task, but it also needs to avoid wasting turns and time along the way.
What Hark’s Internal Tests Show
Hark tested Handoff² on two internal benchmarks: Browsing and Alignment. The Browsing benchmark covers 478 tasks across 238 websites. Handoff² completed 82% of those tasks, up from 73% for v1, and took 46% less time per task.
Hark also ran Claude Opus 5.5 (max) in the same harness. Handoff²’s 82% browsing completion rate landed within 5 points of that result, though the comparison still leaves Claude Opus 5.5 (max) ahead by up to 5 points.
The alignment benchmark measures whether the agent follows important limits during a task. It checks that the agent never checks out without approval and asks for clarification only when a request is truly ambiguous. Handoff² scored 93.8%, compared with 90.3% for v1.
There is another public comparison, but it does not include Handoff² yet. On the Online-Mind2Web human-eval leaderboard, the original Handoff scores 97.67%, while Yutori’s Navigator n1.5 scores 97.3%. Hark has not posted an Online-Mind2Web result for Handoff².
Cloud Access, Plans and Limits
Handoff² runs on Hark’s cloud, and each request gets a dedicated virtual computer with its own browser, file system and terminal. The model outputs cursor and keyboard actions aimed at screen coordinates, giving each task a separate working environment.
Users can reach Handoff² through the web, iOS and Android. There is no local or quantized option, and Hark has not announced a public API or weights. Deployment options and specs are also available through the GPULeaderboard.com chrome extension.
Handoff² ships inside Hark Pro. Hark previewed Handoff in August and shipped it in Hark Pro in October. Since then, Hark says the agent has logged over 5.6M browsing sessions and handled over $20M in purchases.
Per the pricing page, Hark Pro is free. Pro² costs $20 a month for 2x usage, while Pro³ costs $100 a month for 10x usage. All features are offered on all plans.
A Bigger Push Toward Task-Specific Agents
Handoff² arrives as businesses look at agents that can complete tasks across websites and software, rather than only answer questions. Gartner predicts that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in January.
That shift also raises questions about control. Gartner named agentic AI oversight its top cybersecurity trend for 2026 and flagged identity registration and policy-driven authorization for machine actors as areas where enterprises are underprepared.
Handoff²’s alignment score addresses a narrow part of that challenge by testing approval and clarification behavior. Its browsing results address another: whether an agent can finish real tasks without spending too much time on each step.
For now, the clearest picture comes from Hark’s own numbers. Handoff² completes more tasks than v1 on Hark’s internal benchmark, reaches 93.8% on its alignment test and cuts task time by 46% in that comparison. Independent testing will determine how well those gains hold outside Hark’s harness.
Based on
- Hark Releases Handoff², a 2x Faster Computer-Use Agent for Hark Pro — marktechpost.com
- Talkdesk automates long-horizon customer journeys from first signal to final outcome | Markets Insider — markets.businessinsider.com
- Plentisoft | Markets Insider — markets.businessinsider.com
- YouArt Launches YouArt Desktop App, AI Creative Teammate on the Computer | Currency News | Financial and Business News | Markets Insider — markets.businessinsider.com
- AI Orchestration Tools & Agentic Workflows | VentureBeat — venturebeat.com



