Hark Pro: What Brett Adcock Actually Shipped

Hark Pro went live on October 6, 2026. Within forty-eight hours the summary circulating across content sites had three things wrong, and the thing it got right was the thing that matters least.

Let's clear the deck first, because the wrong version is load-bearing for a lot of bad analysis.

Hark is not a Figure product. Brett Adcock founded Figure AI in May 2022 and is still its chief executive. He founded Hark separately in 2025 and self-funded it. Those are two companies that share a founder, and no reporting describes a corporate or financial link between them beyond Adcock himself. Every "robot company builds an app so its robots can learn your laundry habits" take is built on a premise that does not exist.

It does not cost $100 a month. Hark Pro is free on web, iOS and Android. There are paid tiers at $20 and $100 per month, and what they buy is usage headroom — roughly 2x and 10x — not features. The $100 number got pulled out and run as the headline price, which inverts the actual product strategy.

There is no "free trial month." Adcock posted on October 4 that the first 100,000 sign-ups get a paid plan free. That is a land-grab, not a trial funnel, and it tells you the company is optimizing for seats and behavioral data rather than near-term subscription revenue.

Now the parts worth reading.

Hark raised $700 million at a $6 billion valuation in May 2026, roughly two months out of stealth. Adcock put in $100 million of his own money before that round closed.

Look at who else is on the cap table: Nvidia, AMD Ventures, Intel Capital, Qualcomm Ventures. That is not a software-as-a-service investor syndicate. That is four chipmakers, and chipmakers write checks into companies that will consume silicon — in data centers, and in devices.

The company has secured gigawatt-scale Nvidia Vera Rubin capacity and is building its own foundation models. First devices are due in 2027, with AT&T supplying cellular service. Head of design Abidur Chowdhury — formerly at Apple, credited on the iPhone Air, departed late 2025 — has described the devices as portable and useful at home or work, and has explicitly ruled out camera glasses and wearable pins.

Read that together and the app stops looking like the product. A free assistant that ingests your email, calendar, files and payment methods for twelve to eighteen months is how you arrive at a 2027 hardware launch with a user base that already has a reason to buy. The subscription tiers are there to cover inference costs on the heavy users, not to build a business.

Underneath Hark Pro is Handoff, a computer-use agent Hark previewed on August 5, 2026. Its design choices are the most instructive thing about the whole launch, and they cut against how most teams are building agents right now.

It drives a cursor and keyboard, not integrations. Each request gets a dedicated virtual computer with a browser, a file system and a terminal. The agent operates that machine the way a person would, rather than calling a per-service application programming interface (API) for each site it needs to touch. Hark claims it can log into "millions of sites" and run up to six browsers concurrently in the cloud. (Several outlets initially reported 36 concurrent browsers; The Next Web corrected that figure to six.)

The model is purpose-trained, not a general frontier model with a prompt. Hark describes six months of work on computer use specifically: supervised fine-tuning plus reinforcement learning, now moving into mid-training, with pre-training planned later in 2026. That is a company deciding that computer use is its own capability surface and worth owning outright.

The agent works where you can see it. Hark surfaces a small window showing the browser navigating. Most assistants do this invisibly. Showing it is a trust decision, and it is also a hedge — when an agent is about to pay a bill with your card, a visible cursor is the difference between oversight and faith.

State is explicit and inspectable. Persistent memory spans people, places, projects and voice conversations, and users can reportedly inspect or delete what is stored. A Markdown "soul" document appears to govern the assistant's personality. Browser skills are reusable — you define a procedure for a specific site once and it persists.

Anyone who has built a production agent system will recognize that last cluster immediately. A declarative personality file, inspectable long-term memory, and reusable per-site procedures are the same primitives that show up in every serious agent stack, including ours. Hark's contribution is shipping them to consumers with the plumbing hidden.

Hark's headline claim is 97.7% on Online-Mind2Web, across 300 tasks on 136 live websites — 100% easy, 95% medium, 100% hard, evaluated August 4. That number sits on the official leaderboard and the human evaluation was run by Careerflow.ai, not by Hark. It edges Yutori's Navigator n1.5 at 97.3% and clears the ACT-2 system on GPT-5.4 at 92.7%.

That one is real and independently evaluated. The rest deserve more skepticism.

On WebTailBench v2, run on Hark's own harness, Handoff scored 68.6 — behind GPT-5.5 at 72.3, ahead of Claude Opus 4.8 at 66.9. On a Hark internal evaluation judged by a large language model (LLM), Handoff scored 83.2 against Claude Opus 4.8 at 80.5 and GPT-5.5 at 75.2.

So: it wins decisively on the independent benchmark, loses to GPT-5.5 on a self-run benchmark, and wins on a self-run benchmark with a self-selected judge. That is a normal and honest-ish picture of a strong model, and it is worth stating plainly rather than repeating the 97.7% alone. The practical weakness reported from hands-on demos is speed — a DoorDash order took several minutes.

Once granted access to email, calendar, files and payment cards, Hark Pro orders groceries, pays bills, books rides, plans trips and builds documents and web pages. Action Buttons execute single operations like paying a bill or cancelling a service.

The proactive layer is where it differentiates. A feed of action cards proposes things before you ask: approve an expense, reply to an email, book the hotel for the flight you already booked. In demos it caught a California Department of Motor Vehicles (DMV) registration renewal and a low school lunch card balance and offered to handle both. Customizable panels function as mini-dashboards, and daily briefs are designed to decide whether something warrants your attention or to stay quiet.

Staying quiet is a harder product problem than it sounds, and it is the one every proactive assistant has failed so far.

The honest version of the privacy discussion is not "be careful, it wants your data." It is that the product does not function without deep access, and the company knows it.

Chowdhury's position is that Hark sells neither ads nor user data, drawing an explicit contrast with Meta and OpenAI. Card details and passwords sit in an encrypted vault that Hark says it cannot read. Those are meaningful commitments, and they are also commitments rather than verified properties.

Two things are worth watching. First, an agent authorized to transact on your behalf runs into Europe's strong customer authentication rules, which grant no exemption for payments approved by software — Hark has not said whether Hark Pro is available in Europe at all. Second, the business model is unproven. A company with no ad revenue, no data revenue, gigawatt-scale compute commitments and a hardware program is carrying costs that the $20 and $100 tiers plainly do not cover today. Privacy positions are easiest to hold when they are not expensive.

Three things.

The frontier-model-plus-prompt approach is no longer the default for computer use. A funded team spent six months and a research program concluding that the capability is worth training for directly. If browser automation is core to what you ship, that is a signal about where the quality ceiling sits.

Driving a browser and calling integrations are not competing strategies, they are different risk profiles. Integrations are faster, cheaper and more reliable where they exist, and they exist for a tiny fraction of the web. Cursor-and-keyboard control works everywhere and breaks unpredictably. The interesting systems will route between them per task, not pick one.

And proactivity is a product problem before it is a model problem. Hark's architecture for it — inspectable memory, a declarative personality file, reusable site procedures, visible execution — is more copyable than its model. Those primitives are available to anyone building today.

Hark's real test arrives in 2027 with hardware nobody outside the company has seen. The app is the on-ramp.

Sources: The Next Web, TestingCatalog, The Outpost, RuntimeWire, ExplainX, Wikipedia.

Figures reflect reporting as of October 8, 2026. Pricing, availability and feature details on a product this new should be treated as changeable.