{"slug": "hark-handoff", "title": "Hark Handoff", "summary": "Hark introduced Handoff, a computer use agent designed to autonomously navigate websites and complete end-to-end tasks, in its first research preview. The company reports that Handoff outperforms GPT 5.4 by 8 points and Opus 4.8 by 2 points on average across browser-use benchmarks, and it claims top performance on the Online-Mind2Web human evaluation leaderboard. Hark's survey of nearly 3 million screen minutes found that 74.9% of user time is spent in a browser, motivating the focus on web automation.", "body_md": "# Introducing Hark Handoff\n\n## Introducing Hark Handoff\n\nThe world is full of AI assistants, but you would never hire an assistant who couldn’t use a computer.\n\nAt Hark, we’re creating the world’s most capable, personal intelligence - a system to take on tasks and create space to focus on what we choose. And toward that goal, we’ve become the best in the world at building AI that can use the internet for us.\n\nThese are known in the industry as Computer Use Agents, or CUAs. We call ours “Handoff.” Because you just hand off what you don’t want to deal with. And we’re sharing more about Handoff today in our first ever research preview.\n\n## Why computer use on the internet?\n\nWe surveyed users across nearly 3 million screen minutes and found that 74.9% of their time was spent in a browser. People around the world spend 6+ hours a day on the internet, recording trillions of website visits per year across ~300 million distinct web domains. If you're going to automate computer work, the browser is where it happens.\n\nFewer than one in a thousand websites offer a public API, and even those expose only a fraction of what a person can do on the site. Many frequently used sites such as DoorDash, Target, Walmart, OpenTable, Zillow, and LinkedIn have no consumer APIs. To truly unlock all corners of the internet for our personal intelligence, we must build a computer use agent that can access the internet via a browser like a human does.\n\nBuilding a computer use agent that actually works is extremely challenging, and it’s in many ways analogous to building a robot. Just as a robot must perceive and act within an unpredictable physical environment, a web agent must reason and act within an unpredictable digital one.\n\nUnlike a controlled operating system, the internet is a hostile place. Bot blocking, pop-ups, banners, and ads exist specifically to stop, slow, or limit automated agents. Every site is different. Every session introduces new friction. Navigating a browser requires thinking adaptively through dynamic interfaces, and making decisions without a clean, stable world model to rely on.\n\nIn this post, we focus on the model's ability to traverse the open web and complete real, end-to-end tasks.\n\n## Hark Handoff\n\nHandoff is a computer use agent built for long-horizon, autonomous work on the internet. You can ask it to navigate any website, and it handles the task end-to-end, the way a human does.\n\nFor every request, Hark spins up a dedicated virtual computer with its own browser, file system, and terminal. Handoff isn't a script running against a fixed set of sites; it’s an agent that runs in a fully capable environment that can handle the open-ended, unpredictable nature of the web. You can also connect your existing accounts, and Handoff logs in safely on your behalf, so it can act with the full context of who you are - complete with your saved addresses, preferences, and history.\n\nBuilding a computer use agent with even basic functionality requires that it learn to navigate constantly shifting GUIs while absorbing their friction. At Hark, our modeling and engineering teams work hand-in-hand, co-evolving the model alongside the systems that handle the messy reality of the live internet. The result is an agent that can navigate any website - clicking, scrolling, and typing its way through real-time decisions.\n\nHandoff is built for the kinds of tasks that eat up your day; the following examples are all live screen recordings of Hark Handoff in action.\n\n## Model Results\n\nHandoff demonstrates frontier-level performance across three browser-use benchmarks, claiming the top spot on the Online-Mind2Web human evaluation leaderboard and on average outperforming GPT 5.4 by 8 points and Opus 4.8 by 2 points. Handoff also holds a wide margin over Gemini-3.5 Flash and Gemini-2.5 Pro. WebTailBench and Internal evaluations are performed by plugging all models into our internal Hark designed harness.\n\n## Training the Model\n\nOver the past six months, we have been building our understanding of how to train computer use agents. Scaled training data for internet-based computer use is scarce, which makes developing our own training capability a long-term strategic priority for Hark.\n\nWe began with post-training, and the results we are sharing today reflect that work. We have moved into mid-training since, with pre-training planned for later this year. This sequencing is deliberate. Starting at the post-training stage lets our AI organization at Hark move quickly - refining our data pipelines, evals, infrastructure, and training techniques while iterating toward model improvements on the shortest path to results. Each stage builds the foundation for the next.\n\nHandoff’s output is a sequence of cursor and keyboard inputs that directly influence the state of a live browser - targeting x and y coordinates on the screen, while clicking, typing, and scrolling to complete tasks.\n\nGiven the web is so noisy and unpredictable, supervised data alone cannot fully teach agents how to deal with this environment and recover from failures. We believe reinforcement learning (RL) is critical to solve general-purpose computer use. So we use RL to let the agent learn directly from its own experience - succeeding and failing on real tasks, then improving from the outcome.\n\nThe chart below shows pass@1 across our three benchmarks at each stage of post-training: the base model, after SFT, and after RL. Both stages contribute, and their gains stack, making the fully post-trained model strongest across all three benchmarks. Their relative contributions vary by task difficulty. On benchmarks where the base model is already strong (Online-Mind2Web), SFT adds little and RL supplies most of the remaining headroom. On harder distributions, SFT does more of the heavy lifting, with RL adding a further boost on top.\n\n## Cost and Latency\n\nTo be useful in practice, it is not enough for a computer-use agent to be capable. It needs to be efficient too. The efficiency of the model powering Handoff is reflected in both the pricing as well as per-step model latency. Handoff offers superior performance at an order of magnitude lower cost per token than competing frontier models.\n\n## How To Use Handoff\n\nHandoff has exceeded our expectations, and it’s only getting faster and more capable.\n\nComputers were the defining breakthrough of the last century, but they also tethered us to screens and quietly increased our mental load under the guise of convenience. The next step is building computers that can use computers for us. That is the foundation of a more natural interface: one where you speak or type a request, and the work simply gets done.\n\nThat is our vision at Hark. A world where your devices stop consuming your time and start returning it. The Hark platform will be available by the end of summer. We can't wait for you to try it.\n\nSign up to be the first to access Hark below:\n\n## Sign up for Hark beta\n\nWe are reviewing applications to join our platform, and look forward to welcoming you soon.\n\n## You’re signed up\n\nWe are reviewing applications to join our platform, and look forward to welcoming you soon.", "url": "https://wpnews.pro/news/hark-handoff", "canonical_source": "https://hark.com/articles/introducing-hark-handoff", "published_at": "2026-08-05 15:26:18+00:00", "updated_at": "2026-08-05 15:37:08.649170+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-research"], "entities": ["Hark", "Handoff", "GPT 5.4", "Opus 4.8", "Gemini-3.5 Flash", "Gemini-2.5 Pro", "Online-Mind2Web", "WebTailBench"], "alternates": {"html": "https://wpnews.pro/news/hark-handoff", "markdown": "https://wpnews.pro/news/hark-handoff.md", "text": "https://wpnews.pro/news/hark-handoff.txt", "jsonld": "https://wpnews.pro/news/hark-handoff.jsonld"}}