Hark's Handoff² Cuts Browser Agent Runtime in Half on 478 Tasks
Hark's second-generation browser agent completes 82% of internal tasks at roughly half the latency, closing the accuracy gap with Claude Opus 5.5.
Takeaways
- Hark released Handoff², its second-gen browser-use agent, averaging 2x faster end-to-end.
- Task completion rose from 73% to 82% on an internal 478-task, 238-site benchmark.
- Within 5 points of Claude Opus 5.5 max accuracy while running at half the latency.
- Trained with RL that penalizes extra turns, thinking tokens, and wall-clock latency.
- Flight booking is 3.8x faster, shopping 2.5x, long research tasks over 5x.
- Available now to all Hark Pro users; all numbers self-reported.
Hark has released Handoff², an updated computer-use agent for its Hark Pro assistant. The company reports higher task completion, faster browser execution, and more reliable decisions about when to request user input.
Browser agents accumulate latency with every screenshot, model call, network hop, and page load. Hark reduced those cycles by training Handoff² to group related actions into a single response, such as filling several form fields or issuing a sequence of clicks.
The speedup, by workload
On Hark’s internal benchmark of 478 tasks across 238 websites, task completion rose from 73% to 82%. Hark says model execution is 46% faster than the first version, while model, harness, and backend improvements together doubled average end-to-end speed.
| Task or metric | Handoff v1 | Handoff² |
|---|---|---|
| Task completion | 73% | 82% |
| Instruction following and user check-ins | 90.3% | 93.8% |
| Buying coffee | Baseline | 1.7× faster |
| Shopping | Baseline | 2.5× faster |
| Booking flights | Baseline | Up to 3.8× faster |
| Long research tasks | Baseline | More than 5× faster |
Hark gives two examples of the elapsed-time reduction: a 10-minute task falling to about three minutes, and a 45-minute task finishing in roughly 10 minutes. The size of the gain varies with task length and the number of browser interactions required.
Speed comes from fewer browser loops
Handoff² can emit several browser actions in one model response. Its execution harness then carries out those actions together, reducing the number of full request-response cycles needed to complete a task. Each saved cycle removes model inference, network latency, screenshot processing, and time spent waiting for the page to settle.
Hark used reinforcement learning penalties for excessive browser turns, internal reasoning tokens, and latency. The resulting policy batches actions more aggressively. A form with three fields, for example, can be completed through one model response instead of three separate screen interactions.
Those savings compound during long workflows such as comparing flights, completing checkout, or collecting information from several websites. The approach also explains why the largest reported gains appear in research and travel tasks with many sequential steps.
A visual agent in its own computer
Handoff² runs each request inside a dedicated virtual computer containing a browser, file system, and terminal. The model reads rendered pages and produces cursor and keyboard actions aimed at screen coordinates, allowing it to operate websites that lack public developer APIs.
Hark previewed the original Handoff in August and added it to Hark Pro in October. The architecture exposes the agent to the same delays a person encounters on JavaScript-heavy websites, including page rendering, navigation, and asynchronous interface updates. Reducing model turns therefore improves the full workflow even when website response times remain unchanged.
The benchmark has limits
Hark says Handoff² finished within five percentage points of Claude Opus 5.5 at its Max setting while completing tasks in half the time. Both models ran through the Handoff² harness, which controls how actions are represented, grouped, and executed.
The original Handoff model ran through its earlier harness. The v1-to-v2 comparison therefore combines changes to the model, action harness, and backend infrastructure. It does not isolate the contribution of any single component.
All results are self-reported. Hark has not cited an independent evaluation, and the supplied figures omit task-level pass criteria, hardware configuration, run-to-run variance, and failure categories. Those details would be required to reproduce the benchmark or compare its results with other computer-use evaluations.
Founder Brett Adcock separately claimed an 80% improvement in CAPTCHA solving and a 10% gain on end-to-end browser tasks, describing Handoff² as faster and more capable than Claude Opus 5.5 High. Hark did not specify the baseline, metric definition, or evaluation set for those two claims.
Where Handoff² fits
Hark positions the agent for repeatable browser errands and account-based workflows. Supported examples include:
- Ordering food through DoorDash or Uber Eats
- Shopping and checking out on Target or Costco
- Booking restaurant reservations through OpenTable or Resy
- Messaging and recruiting through LinkedIn
- Reserving flights and hotels
- Researching information across multiple websites
Hark says Handoff² is available across all plans and serves as the default computer-use agent inside Hark Pro. The company offers no separate SDK or model endpoint, so developers cannot currently embed the agent directly into their own applications.
Account-connected deployments also require clear policies for credential handling, purchase confirmation, message approval, session isolation, audit logs, and data retention. Hark reports improved judgment about when to check in with users, but the published benchmark score does not define those operational safeguards.
The release combines a more efficient action policy with harness and infrastructure changes. Hark’s reported results suggest that batching actions and reducing model turns can lower computer-use latency substantially, although independent testing is still needed to establish how those gains hold up across sites and workloads.