Hark rolls out Handoff², a second-generation browser agent it claims is twice as fast

Hark has started rolling out Handoff², the second generation of the computer-use agent that powers its personal assistant. According to the company, the new model finishes more tasks and runs end to end at twice the speed of its predecessor, and it says much of that gain comes from reinforcement learning that penalizes unnecessary browser turns, reasoning tokens, and latency.

Hark is the consumer AI lab of Brett Adcock and is separate from Figure. Its Hark Pro application opens a cloud browser and carries out web tasks for the user, such as shopping, booking travel, and managing subscriptions.

All three plans carry the same features, and the Handoff² rollout covers every tier.

Long jobs show the largest gains

On Hark's internal benchmark, Handoff² completed 82% of the suite, up from 73% for the first version. The benchmark contains 478 browsing tasks spread across 238 websites. The new agent also took 46% less time per task, and Hark attributes its rounded twofold end-to-end speed claim to a combination of agent improvements and backend infrastructure changes.

Hark reports that longer workflows benefit most because saved model turns accumulate over the course of a task: some jobs that took 10 minutes now finish in three minutes, and some 45-minute jobs take about 10 minutes.

Fewer turns drive the speedup

Hark's execution harness lets a single model response carry several browser actions, such as filling multiple form fields or performing a sequence of clicks. A model turn is one request-and-response cycle, so batching actions cuts down on network calls, model inference, and waiting on the browser.

The reinforcement-learning objective rewards task success while assigning explicit costs to excessive turns, internal reasoning tokens, and wall-clock latency. During training, the agent also learned to identify repeated action patterns, package them into composite actions, and reuse them later in the same task. Hark describes these temporary composites as resembling macros built from the current browsing context, and says that actions per turn increased while total turns per task declined.

The Opus comparison needs context

According to Hark, Handoff² finished within five percentage points of Claude Opus 5.5 at its "max" setting while completing tasks twice as fast. Both models ran through the Handoff² harness, which controls how actions are represented, grouped, and executed. Because the first Handoff model ran in its original harness, the version-to-version results combine model, harness, and infrastructure changes.

The published comparison comes from Hark's internal suite and omits several details needed for independent replication: the full task set, scoring rules, the exact Opus result, the latency distribution, hardware configuration, and run-to-run variance. As presented, the figures support a product-level comparison inside Hark's environment rather than a general ranking across computer-use systems.

Guardrails improve by Hark's measure

A separate behavioral benchmark tests whether the agent requests approval before checkout and asks clarifying questions only when a request is genuinely ambiguous. Handoff² scored 93.8%, compared with 90.3% for the first version. These checks address practical risks for an agent that can access accounts and initiate purchases, although Hark has not published the benchmark or category-level results.

Millions of sessions, open questions

Since Handoff launched in October 2026, Hark says it has processed more than 5.6 million browsing sessions and over $20 million in purchase volume. Reported tasks include ordering groceries, booking flights, scheduling dental appointments, reserving restaurants, canceling subscriptions, claiming refunds, completing warranty forms, and applying for jobs, activity the company equates to roughly 37 cumulative years of browser time.

Those totals establish usage scale but provide limited evidence about field performance. Hark has yet to publish real-world completion rates, median and tail latency, human-intervention rates, or failure categories. The $20 million figure represents purchases handled by the agent, not Hark revenue.

The source's recipe for faster agents

The report closes with a practical recipe for faster browser agents. Such agents accumulate latency across sequential model calls, page loads, and interface updates; batching safe actions removes some of those round trips, while explicit efficiency penalties keep the training objective from rewarding success without accounting for time and compute. The largest gains should appear in long workflows with repeated actions, matching Hark's reported task breakdown. Its recommendations:

• Action interface: allow multiple clicks, keystrokes, and form updates in one model response, with deterministic execution and per-action results.
• Reward design: score task completion alongside turn count, reasoning-token use, and wall-clock latency, with penalty weights that preserve completion quality and safe behavior.
• Evaluation: run competing models through the same harness and infrastructure, then report completion, token use, and median and tail latency.
• Guardrails: keep approval gates around purchases, submissions, cancellations, and other irreversible actions, even when the agent batches surrounding steps.