Hark's Handoff² Doubles Browser Agent Speed and Beats Claude Opus
Hark's Handoff² computer-use agent completes web tasks twice as fast as the original while closing the accuracy gap with frontier models to within 5 points.
- Hark released Handoff², its second-gen computer-use agent, 2x faster than v1 end-to-end.
- Task completion jumped from 73% to 82% on an internal 478-task, 238-site browsing benchmark.
- Reaches within 5 points of Claude Opus 5.5 accuracy while running at half the latency.
- Trained with RL penalties on turn count, thinking tokens, and latency, enabling learned action batching.
- Rolling out today to free Hark Pro, Pro² ($20/mo) and Pro³ ($100/mo) tiers.
- Flights 3.8x faster, shopping 2.5x faster, long research tasks over 5x faster.
Hark has begun rolling out Handoff², the second generation of the computer-use agent behind its personal assistant. The company reports higher task completion and twice the end-to-end speed of its first model, with much of the gain coming from reinforcement learning that penalizes unnecessary browser turns, reasoning tokens, and latency.
Hark is Brett Adcock’s consumer AI lab, separate from Figure. Its Hark Pro app launches a cloud browser and completes web tasks such as shopping, booking travel, and managing subscriptions.
| Plan | Monthly price | Usage allowance |
|---|---|---|
| Hark Pro | Free | Capped base allowance |
| Hark Pro² | $20 | Twice the base allowance |
| Hark Pro³ | $100 | Ten times the base allowance |
Every plan includes the same features, and the Handoff² rollout covers all three tiers.
Long jobs show the largest gains
Hark reports that Handoff² completed 82% of its internal benchmark, up from 73% for the first version. The suite contains 478 browsing tasks across 238 websites. Handoff² also took 46% less time per task, and Hark attributes its rounded twofold end-to-end speed claim to the agent improvements plus backend infrastructure changes.
| Task | Speedup |
|---|---|
| Buying coffee | 1.7× |
| Shopping | 2.5× |
| Booking flights | 3.8× |
| Long research tasks | More than 5× |
Longer workflows produced the largest reported gains because saved model turns accumulate across the task. Hark says some 10-minute jobs now finish in three minutes, while some 45-minute jobs take about 10 minutes.
Fewer turns drive the speedup
Hark’s execution harness allows one model response to contain 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 reduces network calls, model inference, and browser waiting time.
The reinforcement-learning objective combines task success with explicit costs for excessive turns, internal reasoning tokens, and wall-clock latency. During training, the agent also learned to identify repeated action patterns, package them as composite actions, and reuse them later in the same task. These temporary composites resemble macros created from the current browsing context. Hark says actions per turn increased as total turns per task declined.
The Opus gap needs context
Handoff² finished within five percentage points of Claude Opus 5.5 at its “max” setting and completed tasks twice as fast, according to Hark. Both models ran through the Handoff² harness, which controls how actions are represented, grouped, and executed. The first Handoff model ran in its original harness, so 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, including the full task set, scoring rules, exact Opus result, latency distribution, hardware configuration, and run-to-run variance. 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
Hark’s 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 leave open questions
Since Handoff launched in October, 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. The company equates the activity 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 rather than Hark revenue.
A practical recipe for faster agents
Browser 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.
- 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. Penalty weights should 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.
Comments
