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

Hark has released Handoff², the second iteration of its computer-use agent for Hark Pro, claiming a doubling of speed over the original…

By Vane November 16, 2026 3 min read

Hark has released Handoff², the second iteration of its computer-use agent for Hark Pro, claiming a doubling of speed over the original version.

The company attributes this performance jump to reinforcement learning techniques that penalised wasted turns and latency during training. For teams developing browser agents, this update serves as a case study in prioritising speed alongside accuracy.

Technical specifications and deployment

The agent runs exclusively on Hark’s cloud infrastructure. Each request is assigned a dedicated virtual computer equipped with a browser, file system, and terminal.

Users access the service via web, iOS, and Android interfaces. There is no option to run the software locally or use quantised models.

Performance figures are as follows:

  • Task completion: 82% on Hark’s internal benchmark of 478 tasks, compared to 73% for version 1.
  • Alignment: 93.8% on the internal alignment benchmark, an increase from 90.3% in version 1.
  • Speed gains: Average end-to-end speed is about 2x faster. Specific tasks show higher variance, with flight booking up to 3.8x faster and long research tasks more than 5x faster.
  • External comparison: In the same harness, Handoff² trails Claude Opus 5.5 by up to 5 points on browsing completion. The original Handoff version scored 97.67% on the Online-Mind2Web human-eval leaderboard, though Hark has not yet posted a result for Handoff² on that public board.

Every metric for Handoff² is self-reported. No third-party scores currently exist.

How the speed increase works

Hark co-evolved the agent harness and the reinforcement learning training recipe. The new harness allows the model to emit more actions per turn.

Training specifically penalised excess turns, unnecessary thinking tokens, and end-to-end latency. The model adopted three main habits as a result:

  • Batching actions, such as filling multiple form fields in a single turn.
  • Grouping repeated action patterns into reusable composite actions.
  • Reusing those composites as the task progresses.

Because these savings stack per turn, the improvements compound on longer, browsing-heavy tasks. A 10-minute task can drop to 3 minutes, while a 45-minute task can reduce to about 10.

Comparison with other web agents

Handoff² is a consumer agent within Hark Pro. It operates websites by clicking, typing, and scrolling on live pages rather than calling APIs.

Released on November 12, 2026, it follows a preview in August and a full launch in October of the same year. The service has logged over 5.6 million browsing sessions and facilitated over $20 million in purchases since shipping.

Users send the agent to book groceries, flights, and dentist appointments, process refunds, complete warranty forms, and apply for jobs.

Competitors include Yutori Navigator n1.5 and Meta Muse. Yutori released its model on May 6, 2026, while Meta launched Muse on September 8, 2026. Both competitors operate via API or secure virtual machines. Pricing for Hark Pro varies from free to $100 a month depending on usage multipliers.

Access and cost

Handoff² is available now in Hark Pro on the web, iOS, and Android. Hark has not announced a public API or model weights.

The pricing structure is as follows:

  • Free: Basic access.
  • Pro²: $20 a month for 2x usage.
  • Pro³: $100 a month for 10x usage.

All features are available on every plan.

What it means

For people using Hark Pro, the practical change is that complex browsing tasks finish noticeably quicker. The agent is more efficient at moving between steps without getting stuck in loops or excessive deliberation. This makes the tool viable for time-sensitive activities like booking travel or conducting deep research, though users should treat the performance numbers as internal claims until independent verification appears.

Scroll to Top