# Hark launches Handoff to automate multi-step tasks across the web > Source: > Published: 2026-08-05 16:09:24+00:00 [Hark](https://hark.com/?ref=runtimewire) founder [Brett Adcock (@brettadcock)](https://x.com/brettadcock?ref=runtimewire) announced Handoff on August 5, 2026, giving Hark its first public product: an agent that gets a virtual computer, opens websites and attempts to finish multi-step jobs for the user. The [research preview, first reported by VentureBeat](https://venturebeat.com/technology/ai-startup-hark-unveils-first-product-an-affordable-fast-computer-use-agent-hark-handoff?ref=runtimewire), is the first concrete look at what Adcock plans to build after Hark raised a $700 million Series A in May 2026 at a $6 billion valuation. [Hark's website](https://hark.com/?ref=runtimewire) says the company is entering beta and reviewing applications, while VentureBeat reported that access to the software platform was planned for later in August. [VentureBeat's launch report](https://venturebeat.com/technology/ai-startup-hark-unveils-first-product-an-affordable-fast-computer-use-agent-hark-handoff?ref=runtimewire) says Handoff can identify and message recruiting candidates on LinkedIn. In Hark's launch video, Adcock said he uses the agent for "all of my recruiting efforts end to end." Adcock entered startups through recruiting. An [Archer Aviation proxy statement](https://s202.q4cdn.com/174276461/files/doc_financials/2022/Archer-2022-Proxy-Statement-%281%29.pdf?ref=runtimewire) says he earned a business administration degree from the University of Florida and worked in investment banking and hedge funds before co-founding the online hiring marketplace Vettery. [Adecco acquired Vettery in 2018](https://www.adeccogroup.com/our-group/media/press-releases/2018-the-adecco-group-announces-acquisition-of-vettery?ref=runtimewire), giving Adcock a foundation for his subsequent bets on electric aircraft at Archer Aviation and humanoid robots at Figure AI. That experience makes recruiting a natural proving ground for Handoff. It is a workflow Adcock knows, full of repetitive browsing and messaging as well as sensitive decisions that can expose the limits of autonomous software. ### A product before the hardware For each request, [Hark Handoff](https://hark.com/articles/introducing-hark-handoff?ref=runtimewire) creates a dedicated virtual computer equipped with a browser, file system and terminal. It clicks, scrolls and types through websites instead of waiting for each service to provide an API. Users can connect existing accounts, allowing the agent to work with saved addresses, preferences, payment methods and account history. VentureBeat reported that Hark showed Handoff ordering dinner, booking flights and messaging job candidates. The publication also cited hotel reservations, online research and shopping among the tasks Hark says the agent can perform. That scope is central to Adcock's bet. [Hark says](https://hark.com/articles/introducing-hark-handoff?ref=runtimewire) an analysis of nearly 3 million minutes of screen activity found that 74.9% of people's time was spent in browsers, while fewer than one in 1,000 websites offered a public API. Those are Hark-produced figures, though the basic product decision is straightforward: a useful general agent has to operate the websites people already use. Handoff also narrows Hark's initial ambitions. Adcock has described Hark as a vertically integrated personal intelligence company spanning models, software and custom devices. The [May financing announcement](https://www.businesswire.com/news/home/20260521171628/en/Hark-Raises-%24700M-Series-A-at-a-%246B-Valuation?ref=runtimewire) said Hark had raised $700 million at a $6 billion post-money valuation, led by Parkway Venture Capital. Nvidia, AMD Ventures, Intel Capital, Qualcomm Ventures, Salesforce Ventures, ARK Invest, Brookfield, Greycroft, Prime Movers Lab, Align Ventures and Tamarack Global participated. RuntimeWire [reported when the round closed](/article/hark-raises-700m-series-a-brett-adcock) that the size of the financing placed unusual pressure on Hark to turn a broad promise about personal AI and native hardware into something users could evaluate. Handoff is Adcock's first answer. It is software, runs in the browser and can reach users before Hark ships a device. ### The benchmark lead has boundaries [Hark says Handoff scored](https://hark.com/articles/introducing-hark-handoff?ref=runtimewire) 97.7 on Online-Mind2Web, a benchmark that tests agents against live websites. Hark compared that result with 92.8 for OpenAI's GPT 5.4, 84.1 for Anthropic's Claude Opus 4.8 and 69 for Google's Gemini 2.5 Pro. The claim remains largely Hark's to prove. The public Online-Mind2Web leaderboard did not independently verify Hark's 97.7 score, and Hark's comparisons excluded newer frontier models for which equivalent public results were unavailable. Hark did not lead every test it disclosed. On WebTailBench v2, VentureBeat reported that Handoff scored 68.6 against 72.3 for GPT 5.5. Hark ran that benchmark and an internal evaluation in its own harness, using an internal model judge to calculate pass rates. The speed figures need similar context. Hark reports 0.8 seconds of model latency per turn and token prices of $0.18 per million input tokens and $2.37 per million output tokens. Hark produced the rival measurements in its own harness, with competing models set to their highest reasoning levels. The cost claim may prove to be Handoff's strongest opening. An agent that loops through dozens of browser actions can become expensive quickly, so lower inference costs could matter as much as a narrow benchmark lead. That advantage depends on Hark delivering the reported performance under ordinary user workloads. ### Hark still depends on an unnamed base model Hark says the preview model was developed through supervised fine-tuning followed by reinforcement learning. Hark acknowledges that it has completed only post-training so far, with pre-training planned for later this year. That sequence means the initial Handoff model sits on top of a base model Hark did not pre-train. Hark has not identified that model or detailed the mix of proprietary and open training data behind Handoff. The omission matters because Hark's broader pitch rests on controlling the models, software and hardware as one system. Handoff can still become a useful product without owning every layer on day one. Starting with post-training lets Hark test its agent architecture, training pipeline and user experience while its larger model program is still being assembled. It also gives Adcock a way to gather real task data before committing Hark's capital to pre-training. ### Account access raises the stakes The same architecture that makes Handoff useful creates its sharpest product risk. An agent capable of logging into airline, shopping and recruiting accounts can see addresses, payment information, messages and files. A mistake can create a purchase, send a message or expose private data rather than merely produce a bad chatbot answer. Hark says security and privacy are priorities, but the technical preview does not yet specify access controls for its virtual computers, how long files and credentials are retained, or who can inspect a session. Those details will shape whether Handoff remains a consumer convenience or becomes software that companies trust with operational work. Adcock has now put a product behind Hark's unusually large financing and valuation. Handoff shows a credible route from his recruiting experience to a broader personal agent, and its virtual-computer design gives the model a practical way to reach websites that were never built for AI. The next test will happen outside Hark's own demos and evaluation harness, when users hand the agent real accounts and expect it to finish the job safely.