Hark makes its Handoff² browser agent twice as fast: task success rises from 73% to 82%
A browser agent that deliberates over every click is slower than doing the job yourself. Brett Adcock's Hark has released Handoff², an agent that buys groceries and books tickets on its own, and it is twice as fast as the first version. The model was trained to fill in all the fields of a form in a single turn. Time per task fell 46%, and the success rate on an internal set of 478 tasks rose from 73% to 82%. So far, the checks are internal only.
AI-processed from TG: @ai_machinelearning_big_data (Machinelearning); edited by Hamidun News
The main problem with browser agents is not stupidity but slowness. An agent that spends a minute thinking over every click saves no time: it is easier to do everything yourself. Hark, the consumer AI lab of Figure AI founder Brett Adcock, has released Handoff², the second generation of its agent that drives a browser on its own and carries out errands: buying groceries, booking tickets, scheduling doctor's appointments.
According to the company, it is on average twice as fast as the first version. Picture a ten-field form: the old agent would fill the fields in one by one, while the new one fills them all at once. At the same time, the share of successfully completed tasks on an internal set of 478 tasks across 238 sites rose from 73% to 82%, so the speedup did not come at the cost of quality.
What got better in Handoff²
The first version of Handoff launched in October. According to Hark, it has since run more than 5.6 million browser sessions and paid for purchases worth more than $20 million. For an agent that people trust with their money, that is a notable scale.
Here is what changed in the second version, according to the company:
- an average speedup of 2x, up to 3.8x for flight booking and more than 5x for long research tasks;
- the share of successfully completed tasks on the internal set of 478 tasks rose from 73% to 82%;
- on an internal behavior test, the score rose from 90.3% to 93.8%. The test checks, for example, that the agent will not complete a purchase without the user's consent;
- according to Adcock, the system solves captchas 80% better.
All figures are internal, and there is no independent verification yet.
How the agent was taught not to waste turns
The key is that the "harness" and the reinforcement learning (RL) were developed together. The harness is the scaffolding around the model: which actions are available to it, what it sees on the page, and how its steps turn into clicks and typed text. Normally a model takes one step per turn: it clicks, waits for the page, thinks, and clicks again. Each such cycle costs seconds.
In Handoff², the model could take several actions in a single turn. To make sure it used that ability, the training added penalties for extra turns, excessive reasoning and overall latency. The model began combining obvious actions on its own, for example filling in several form fields at once.
The second technique is composite actions. When the same sequence of steps keeps recurring in tasks, such as logging in or picking a date in a calendar, the model bundles it into a single action and reuses it. It is similar to a macro, except that the model devises and reuses it by itself.
The result: each task takes 46% less time. Combined with improvements on the server side, that adds up to the 2x speedup. It is most noticeable on long tasks, where the savings on each step accumulate.
What this means for developers and product managers
First: an agent's speed is not just a matter of model power or hardware. A significant part of the gain comes from how the action interface is designed and what the model is rewarded for during training. If you are building your own agent, check whether you are forcing it to take one step at a time where it could take a batch.
Second: success rate and speed improved together. This matters because penalties on reasoning length often hurt quality. Hark says that did not happen here, but this can be checked only on its own task sets.
Third: the separate "don't buy without consent" test shows what the market is watching. When an agent spends real money, the key metric is not only the percentage of completed tasks but also discipline: not doing anything extra. A score of 93.8% means that roughly one case in sixteen still fails the check. For a service that pays for purchases, that is not a small number.
Handoff² is already rolling out to all plans. The next question is whether ordinary users, rather than the company's internal benchmarks, will confirm its speed and reliability.
- Nov 18, 2026Zeroing Out Less Than 0.05% of Weights Taught a Large LLM New Skills as Well as Fine-Tuning
- Nov 18, 2026MoEngage Plugs AI Agents Into Its Existing APIs: Glance's Push CTR Up 38%
- Nov 18, 2026Tencent Hy4 flash scores 2.79 out of 4 vs. 2.81 for Haiku 5.5 and costs 18 times less per task
- Nov 18, 2026AI agent left SOC 27 minutes to respond versus 53 for a junior attacker
Key Questions Answered
What got better in Handoff²?
The first version of Handoff launched in October. According to Hark, it has since run more than 5.6 million browser sessions and paid for purchases worth more than $20 million. For an agent that people trust with their money, that is a notable scale.
How the agent was taught not to waste turns?
The key is that the "harness" and the reinforcement learning (RL) were developed together. The harness is the scaffolding around the model: which actions are available to it, what it sees on the page, and how its steps turn into clicks and typed text. Normally a model takes one step per turn: it clicks, waits for the page, thinks, and clicks again. Each such cycle costs seconds. In Handoff², the model could take several actions in a single turn.
What this means for developers and product managers?
First: an agent's speed is not just a matter of model power or hardware. A significant part of the gain comes from how the action interface is designed and what the model is rewarded for during training. If you are building your own agent, check whether you are forcing it to take one step at a time where it could take a batch. Second: success rate and speed improved together. This matters because penalties on reasoning length often hurt quality.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.