The world’s most-funded computer use agent startup launched its first product today. Hark, the AI lab founded in late 2025 by serial entrepreneur Brett Adcock, released Hark Handoff on August 5, 2026: a browser-based agent that can navigate any website autonomously, without needing an official API. Sign-ups opened at hark.com, with the full platform available later this month.
The launch follows Hark’s $700 million Series A round in May 2026 at a $6 billion valuation, backed by NVIDIA, AMD, Intel Capital, Qualcomm Ventures, and Salesforce Ventures. Three months from funding close to first product is a fast timeline, and the early benchmark numbers are aggressive.
Why the Browser Became the Real Enterprise Automation Layer
Before evaluating whether Handoff works, the framing matters. Hark surveyed users across nearly 3 million screen minutes and found that 74.9% of screen time is spent in a browser. People average more than 6 hours per day on the internet, generating trillions of site visits annually.
More importantly: fewer than one in a thousand websites expose a public API, and even those cover only a fraction of what a logged-in human user can do. Sites including DoorDash, Target, Walmart, OpenTable, LinkedIn, and Zillow have no consumer-facing APIs at all. Standard enterprise automation tools (Zapier, Make, traditional RPA) require connectors or stable API endpoints to function. When those connectors do not exist, the workflow cannot be automated.
Hark’s case is that the browser is where enterprise automation has to happen if it is going to cover the full scope of digital work. And to automate the browser reliably, you need a model trained to act, not just to predict text.
How Handoff Works
For every task, Handoff spins up a dedicated virtual computer with its own browser, file system, and terminal. The agent connects to a user’s existing accounts and then logs in on their behalf, with access to saved addresses, payment methods, and preferences. It reads the visual structure of each page and makes click-by-click, keystroke-by-keystroke decisions in real time.
Hark describes the challenge bluntly: “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.” The training pipeline used supervised fine-tuning followed by asynchronous reinforcement learning with the GRPO algorithm, according to materials shared with VentureBeat ahead of launch.
The architectural distinction Hark draws versus standard LLMs: Handoff predicts the next action (a specific click or keyboard input at a specific screen location) rather than the next token. Whether this is a meaningful distinction or a framing choice will depend on independent benchmarking.
Adcock demonstrated the product himself in a launch video, verbally requesting a floral bouquet and watching Handoff navigate a florist’s site to place the order. He also said he uses Handoff for “all of my recruiting efforts end to end,” including outreach and interview scheduling on LinkedIn.
Benchmark Results
Hark published performance data across three browser-use benchmarks in its research preview.
| Benchmark | Hark Handoff | vs GPT-5.4 | vs Claude Opus 4.8 |
|---|---|---|---|
| Online-Mind2Web (human eval) | Top ranked | +8 points | +2 points |
| WebTailBench | Frontier-tier | Outperforms | Outperforms |
| Internal Hark harness | Frontier-tier | Outperforms | Outperforms |
Online-Mind2Web scores are taken from the official leaderboard where available. Numbers marked with an asterisk in Hark’s release are self-reported. Handoff also shows a wide margin over Gemini-3.5 Flash and Gemini-2.5 Pro across the same benchmarks.
Independent replication has not been published. Hark’s harness for WebTailBench and the internal evaluation is designed by Hark itself. Enterprise teams evaluating Handoff should treat these benchmark gaps as a signal worth investigating, not a substitute for running their own tests on their specific task distribution.
On cost, Hark claims Handoff is faster and cheaper than GPT-5.5 and Opus 4.8 for browser-based tasks. No specific pricing was published on launch day.
The Team and Capital Behind It
Brett Adcock is now running two companies simultaneously: Figure AI (humanoid robotics, raised $675M in partnership with OpenAI) and Hark. He seeded Hark with $100 million of his own capital when he founded it in late 2025. By March 2026, the company had 70 employees and was already running a cluster of NVIDIA B200 GPUs.
The design lead is Abidur Chowdhury, formerly an industrial designer at Apple credited with the iPhone Air. Hark’s engineering team includes researchers with backgrounds in large-model training, tokenization, and safety, plus hardware engineers from Apple, Tesla, and Meta.
One detail that stands out in the investor list: NVIDIA, AMD, Intel Capital, and Qualcomm Ventures all participated in the round. Chip companies do not typically lead consumer AI product bets. Their involvement suggests they see Handoff as a workload that will consume substantial inference compute at scale, which aligns with the architecture choice of spinning up a dedicated virtual machine per user session.
Hark also confirmed that its models are being trained on Figure’s robot data, though no plans exist to combine the two companies.
What Enterprise Teams Should Know
Hark Handoff enters a category that is filling quickly. Prentis AI raised $100 million in July 2026 for AI agents automating office workflows through computer use, with an emphasis on structured document tasks and desktop applications. OpenAI’s Astra model is building multi-agent coordination that includes computer-use capabilities. Hark’s differentiation is the emphasis on open-web browser navigation for tasks that have no API alternative.
What Handoff is positioned for today:
- Recruiting: sourcing, messaging, and scheduling candidates on LinkedIn without a LinkedIn API partnership
- Procurement: price comparison and checkout flows on retail sites
- Travel: booking flights and hotels across carrier and OTA sites
- Research: synthesizing information from multiple web sources in parallel
- Reservations: restaurant and event booking on OpenTable, Resy, and similar platforms
The questions enterprise teams need to answer before deploying:
Security and data isolation. Each session spins up a virtual machine with access to the user’s credentials. Who can access the files and session data those machines generate? Hark’s position as of August 5: “security and privacy is a primary focus, but this is a technical preview.” Specifics are expected when the platform formally launches later this month.
Legal exposure. Automating interactions on sites that prohibit bot activity in their terms of service creates risk for enterprise deployers. Legal and compliance teams should review the specific sites any Handoff workflow would touch before integrating into production.
Base model opacity. Hark did not disclose the underlying model at launch. The company has completed only post-training so far, with pre-training planned for later in 2026. For enterprises with model provenance requirements, this is a gap.
Benchmark independence. Third-party validation is pending. The two non-leaderboard benchmarks use Hark’s own harness.
The Larger Signal
Hark Handoff’s significance goes beyond whether this specific product delivers on its benchmarks. A $700 million Series A, backed by semiconductor makers and enterprise software investors, says something about where the next enterprise software layer is heading.
The pattern across Prentis AI, Hark, and related plays is consistent: investors believe that agents capable of operating any software through a visual interface will replace a substantial portion of human digital labor within the next few years. Not just internal enterprise applications, where there are established APIs and connectors. The entire open web.
For enterprise AI buyers, the practical implication is that the scope of what can be automated without custom engineering is about to expand significantly. Tasks that require a human today because no API exists, no Zapier connector has been built, or the workflow is too dynamic for traditional RPA, are exactly the tasks this generation of computer use agents is targeting.
The other implication is governance. Computer use agents that act with a user’s credentials on real websites are a different risk surface than a chatbot with a constrained tool set. Frameworks for what these agents are authorized to do, what they are not, how their actions are logged, and how they can be stopped, are things enterprise teams need to work out now, not after deployment.
Enera helps enterprise teams design AI-native workflows, including governance frameworks for agentic deployments across GTM, operations, and knowledge work. If you are evaluating where computer use agents fit in your stack and what oversight they require, book a call to work through the specifics.