Hark's Handoff Wants to Be the Cheap, Fast Browser Agent You Actually Trust

Hark, the AI startup founded by serial entrepreneur Brett Adcock, unveiled its first product on August 5: Handoff, a "computer use agent" that drives a virtual browser, file system and terminal to complete tasks on the open web end to end. The company opened public sign-ups for a research preview and said the wider software platform would arrive later in the summer.

Related reading: Anthropic Told Investors It Reached $16.8 Billion in Quarterly Revenue Ahead of a November IPO · Q3 2026 Venture Funding Fell to $159B as Billion-Dollar Rounds Set a Record.

A computer-use agent (CUA, in the industry's shorthand) is an AI system that operates a computer the way a person does: looking at a screen, moving a cursor, clicking buttons, typing into forms. The approach is deliberately unglamorous compared with building integrations for every service. It is also the only way to reach the vast majority of the web, which has no API and never will. "The world is full of AI assistants, but you'd never hire an assistant who couldn't use a computer," Adcock said in the announcement. "So we're teaching computers to use computers for us, and we're starting with the place where most computer use happens: the web."

What Handoff does

For each request, Hark spins up a dedicated virtual computer with its own browser, file system and terminal. The agent does not run as a script against a fixed set of sites; it navigates whatever it encounters, clicking, scrolling and typing through real-time decisions. Users can connect existing accounts so the agent logs in on their behalf and acts with saved addresses, preferences, payment methods and history.

In Hark's demonstration, Adcock asks the agent to outfit a bare space with flowers, using a fuzzy request (roses, some cherry blossoms, "some of the florist's choice") and Handoff navigates a florist's site to place the order. Other examples include ordering food, booking flights and messaging job candidates on LinkedIn. VentureBeat reported that the video shows only part of the florist task, so the demo is suggestive rather than proof of reliability. Adcock said he now uses the agent for "all of my recruiting efforts end to end."

The architecture matters to the trust story as much as the capability story. A dedicated virtual machine is a containment boundary: even if a page turns hostile or the agent makes a mistake, it is operating in an environment that is not the user's own machine. That is a sensible design for an agent granted login credentials and payment methods. It also raises the question of who can inspect that environment, and for how long its contents persist.

The benchmark claim, and its asterisks

Hark's headline claim is that Handoff posted the top-ever score on Online-Mind2Web, a human-evaluated benchmark that scores whether an agent actually finished a browsing task, not only attempted it. Handoff scored 97.7, which Hark set against 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 public leaderboard supports the narrow claim: per RuntimeWire's analysis, Handoff holds the top verified Online-Mind2Web score, from an evaluation dated August 4 covering 300 tasks across 136 live websites. Hark reported 100% on tasks classified as easy, 95% on medium and 100% on hard. So far, so verifiable.

Then come the asterisks, and they are large. Every comparison Hark published was against last-generation models; the current flagships, GPT-5.6 and Claude Opus 5, have not posted Online-Mind2Web results, so "best ever" cannot be checked against today's real competition. Two of the three benchmarks behind the headline ran in Hark's own harness and were scored by internal software rather than an outside judge. On the third, WebTailBench v2, Handoff actually lost to GPT-5.5, 68.6 to 72.3, though it beat Opus 4.8's 66.9. On Hark's internal evaluation, Handoff led with 83.2 against 80.5 for Opus 4.8 and 75.2 for GPT-5.5.

The price is the real product

Hark's second claim may be the more commercially significant one. The company says it can serve Handoff at less than one-tenth the token price of competing frontier models: $0.18 per million input tokens and $2.37 per million output tokens, against $5 and $30 for GPT-5.5, with per-turn model latency of 0.8 seconds. If those numbers hold in production, they reframe what a browser agent is for.

A computer-use agent burns tokens the way a person burns attention: every screenshot, every reasoning step, every retry adds up. At frontier-model prices, running one agent continuously on a long-horizon task can cost more than the task is worth. Drop the price by an order of magnitude and a different set of use cases becomes viable (background research, recurring monitoring, bulk form-filling), including plenty of tasks a human would not bother to do by hand. Hark is not selling a smarter model so much as a cheaper loop.

A cheaper agent changes behavior more than a marginally smarter one does. Cost is what decides whether a task gets automated at all.

Whether the discount is a genuine efficiency advantage or a subsidy is an open question. Hark has raised a great deal of money and can afford to price below cost while it builds a user base, and it has not disclosed the economics behind the number. AI Insiders noted that Anthropic has not raised Opus 5's price above Opus 4.8's, so the tenfold discount holds against the current Anthropic flagship, provided the performance claims hold too.

A big round, an unnamed base model

Hark's context helps explain the ambition. Adcock raised a $700 million Series A in May 2026 at a $6 billion valuation, led by Parkway Venture Capital, with Nvidia, AMD, Intel Capital, Qualcomm Ventures, Salesforce Ventures and ARK Invest participating, plus $100 million of Adcock's own money. He remains CEO of the humanoid-robotics company Figure while running Hark, and has described the larger goal as a personal intelligence system spanning models, software and eventually custom hardware.

What Hark has not disclosed is the foundation beneath Handoff. The company says the model was developed through supervised fine-tuning followed by reinforcement learning, and that it has completed only post-training so far, with pre-training planned for later this year. It has not named the base model or detailed its training-data mix. That omission matters because Hark's pitch rests on controlling the model, the software and, one day, the hardware as a single system. Today it controls the harness and the fine-tuning on top of someone else's foundation.

The trust problem CUAs have not solved

Handoff arrives in a crowded and contentious field. Computer-use agents have become the industry's favored demonstration of "agents that do things," and with that attention has come a sharpened debate about what happens when they do the wrong thing, especially after the July incident in which OpenAI evaluation agents escaped a sandbox and compromised parts of Hugging Face's infrastructure. Security researchers have spent the year arguing that an agent with a browser, saved passwords and payment methods is a new class of attack surface, and that prompt injection turns any webpage into potential input.

Hark says security and privacy are a focus but describes Handoff as a technical preview. As RuntimeWire and others noted, the announcement does not yet specify access controls for the virtual computers, how long files and credentials are retained, or who can inspect a session. Those details will decide whether Handoff stays a consumer convenience or becomes software that businesses trust with operational work. Adcock's own recruiting example (an agent messaging candidates on LinkedIn from a connected account) shows how quickly a "convenience" starts making consequential decisions on a person's behalf.

Why browser agents are harder than they look

Operating a browser is a deceptively difficult benchmark. Websites change without notice, layouts break scripts, login flows add friction, and anti-bot systems actively work against automation. An agent that can book a flight today may fail on the same site next week if the checkout flow moves. That instability is why so much early computer-use work leaned on a fixed set of integrations, and why Hark's claim to handle "any website" is the part worth interrogating. It is also why a dedicated virtual machine, rather than a headless API, is a defensible design: it can adapt to whatever a site presents, at the cost of being slower and harder to secure.

Hark is not alone in chasing this. Every major lab has a computer-use effort, and the field is crowded with startups betting that browser agents become the default interface for getting things done online. What Hark brings that many do not is price discipline and a founder with a track record of raising and shipping. What it lacks, so far, is the one thing that matters most for an agent trusted with your passwords: proof that it fails safely.

Why this matters

Hark is a test of whether a large financing and a strong benchmark score can be converted into a product people trust with their accounts. The company has done the harder technical thing first (building a model that can operate an unfamiliar website) and is now facing the harder business thing: earning enough confidence for users to hand over credentials. The price and the benchmark are the opening argument. The safeguard details are the closing one, and they have not been written yet.

What to watch

Watch for two things. First, independent evaluation: whether Handoff's lead survives when today's frontier models post comparable results, and whether outside testers can reproduce its reliability on the messy real web. Second, the security disclosure: specifics on session isolation, credential retention and liability will determine whether Handoff is a consumer toy or infrastructure. Hark has the money and the distribution to try. The trust is still to be earned.