Hark Handoff Review: The Computer-Use Agent That Actually Clicks

Hark Handoff is one of the clearest demonstrations yet of a computer-use agent completing ordinary work on the live web. It clicks, types and scrolls through sites it was not given a custom integration for. Hark shows end-to-end food orders, shopping, restaurant reservations, candidate outreach, cross-site research and flight booking.

Our buyer verdict on the research preview is deliberately narrower than the launch story: Handoff is a strong candidate for a controlled pilot when a valuable workflow spans unfamiliar web interfaces and useful APIs do not exist. It is not yet a reason to replace stable APIs, deterministic automation or human approval for consequential actions.

This review owns that product decision. If you only need broad read access for research, see our Agent Reach review. If you are choosing browser infrastructure, read the Lightpanda production review. For a full operating-cost model, use cost per accepted agent action.

What is Hark Handoff?

Hark Handoff is a computer-use agent built to complete long-running tasks on live websites by operating a virtual computer. For each request, Hark says it provisions a dedicated environment with a browser, file system and terminal. The agent then produces cursor and keyboard actions, observes the changed screen and continues until it reaches an outcome.

Hark's official Handoff research preview reports that 74.9% of nearly three million observed screen minutes were spent in a browser and that fewer than one in a thousand websites expose a public API. Those are vendor-reported measurements, not an independent census, but they identify the commercial gap accurately: important work still lives behind interfaces designed for people.

The product name also explains the promise. A user hands off the objective rather than recording a macro or specifying every selector. The model is expected to understand a new page, decide what to do next and recover when the site behaves differently.

How does Hark Handoff work?

1. You describe an outcome. Examples include finding a suitable flight, ordering dinner or researching a shortlist.
2. Hark starts an isolated virtual computer. The environment includes the browser and general computer tools needed for the task.
3. Handoff observes and acts. It targets screen coordinates, clicks, types, scrolls and reads the resulting state.
4. The agent keeps the task state. It reasons across several steps and, in the published demos, across several sites.
5. Connected accounts add personal context. Hark says users can connect accounts so Handoff can use saved addresses, preferences and history.

This resembles the perception, action and feedback loop of a robot more than a conventional API integration. That connection is not accidental. Hark founder Brett Adcock's official biography describes Figure as an effort to give AI a body. Handoff points the same systems problem at unpredictable digital interfaces: perceive the environment, choose an action, inspect the result and adjust.

What can Handoff do today?

The research preview shows six useful task families. Treat them as evidence of breadth, not as a service catalogue or reliability guarantee.

The mundane nature of these examples is the point. A travel portal or supplier form may be too small to justify a bespoke integration and too variable for a brittle macro. A general agent can make the human interface programmatic without asking the site owner to build a new API.

How strong are Hark's benchmark claims?

The results are promising, but buyers should separate the public benchmark from Hark's internal evaluations. Hark reports first place on the human-evaluated Online-Mind2Web leaderboard and says Handoff outperforms named frontier models on average across three browser-use evaluations. WebTailBench and Hark's internal evaluation were run through Hark's own harness, according to the launch article.

The Online-Mind2Web research paper is a useful primary reference. It defines 300 tasks across 136 live websites and reports about 85% agreement between its automatic judge and human judgment. Live sites make the test more realistic than a frozen page set, but they also change between runs. The benchmark measures broad web competency. It does not prove success on your exact accounts, approval rules, data or month-end volume.

What the benchmark does and does not tell a buyer

Ask for pass rates by task family, repeated-run variance, intervention rate, timeout policy, failed-transaction handling and the exact denominator. A leaderboard result belongs in vendor discovery. Your acceptance set belongs in procurement.

Handoff vs API, RPA and browser automation: which should you use?

Use the narrowest reliable interface that solves the business problem. A computer-use agent is valuable because it covers the gaps between systems, not because every supported system should be reduced to pixels and clicks.

A strong production design often uses three lanes. APIs handle deterministic reads and writes. A computer-use agent covers the small set of portals or edge cases without usable interfaces. A human approves money movement, external communication, legal commitments and ambiguous exceptions.

What does Hark Handoff cost?

Hark had not published a general Handoff price, enterprise contract or service-level commitment when we reviewed the public launch material on 7 August 2026. The company says its model has lower per-token cost than competing frontier models, but the public buyer question is total cost per accepted task, not token price.

Include virtual-computer runtime, model turns, browser latency, retries, captchas or blocked sessions, human approval, failed purchases, credential operations, observability and maintenance. A cheap model can produce an expensive workflow if it needs many steps or frequent rescue. Use the worksheet in our AI-agent cost-per-action guide before comparing a vendor quote with internal labor.

Is Hark Handoff safe for business use?

A dedicated virtual computer is a good isolation primitive, not a complete security model. Handoff can encounter hostile page content while holding logged-in access and the ability to click. That combines untrusted input with real authority.

NIST's 2026 agent-hijacking analysis covers more than 250,000 attack attempts against 13 frontier model targets. At least one successful hijacking attack was found against every target. This does not evaluate Handoff specifically. It does show why a browser agent should be designed on the assumption that malicious instructions will eventually reach it.

Controls to require before a real pilot

• Dedicated identities: use a pilot account with the minimum data, balance and permissions needed for one workflow.
• Approval boundaries: pause before checkout, booking, sending a message, accepting terms, downloading an executable or changing account settings.
• Task-scoped secrets: expose credentials only to the relevant site and run, then revoke or rotate them.
• Complete evidence: retain the requested goal, page states, actions, approvals, final receipt and failure reason.
• Network and file controls: restrict destinations, downloads, uploads and terminal commands to the pilot's needs.
• Manual fallback: route uncertainty, blocks and mismatched totals to an operator without losing state.
• Terms and data review: confirm that account automation, retention and model processing are permitted for the selected sites and data.

NIST's agent tool-access taxonomy places browser use in an untrusted environment with constrained write access and general computer use in the broader write category. That distinction is a practical architecture rule: give the agent less authority than the virtual machine technically allows.

Which business workflows are a good fit?

The best first task is tedious, valuable, easy to verify and cheap to reverse. The worst is vague, high-impact and impossible to audit. “Collect three comparable supplier quotes” is a pilot. “Handle procurement” is a liability statement disguised as a prompt.

How should a company pilot Hark Handoff?

1. Choose one bounded workflow. Freeze 30 to 50 representative tasks, including awkward pages, expired sessions and unavailable options.
2. Record the human baseline. Measure completion, elapsed time, review time, error rate and total labor.
3. Define acceptance before the demo. State the correct final outcome, required evidence, forbidden actions and approval points.
4. Use a disposable identity. Start with synthetic or low-value data, narrow access and a hard transaction limit.
5. Run repeated trials. One successful video is not the denominator. Repeat tasks on different days and preserve failures.
6. Compare alternatives. Price the official API, a deterministic script, a hybrid design and the current manual process.
7. Scale only the winning lane. Keep exceptions manual and move stable subflows to APIs when their economics justify it.

Computer-use pilot scorecard

Our 30-60-90 day AI-agent pilot plan provides the governance cadence. The agent eval sandbox checklist covers isolation, secrets, network policy and evidence in more detail.

Should you buy Handoff or build your own computer-use workflow?

Buy or pilot Handoff when general web competence is the scarce capability. Build a hybrid when your advantage lives in workflow rules, proprietary data, approvals and integration with internal systems. Training a general computer-use model is not a normal product backlog. Owning the orchestration around a specialist model is.

Before procurement, ask Hark about regional availability, data processing, retention, account connection, model training, audit export, approval controls, benchmark traces, support, incident handling, concurrency, rate limits, pricing and termination. Then ask your team which controls must remain outside the vendor boundary.

Wavect's AI product engineering service can map that boundary and build the API, policy and human-review layers around an agent pilot. The Twinsoft AI case study shows how we approach a production AI system, while the prototype-to-production decision guide helps scope the hardening work. If the workflow is specific, request an architecture review.

Research boundary

Reviewed on 7 August 2026 from Hark's launch article and public site, Brett Adcock's official biography, the Online-Mind2Web paper and NIST agent-security material. We did not receive paid access, connect a personal account or run an independent Handoff benchmark. Hark's product claims are identified as vendor-reported. Confirm current availability, pricing, controls and contractual terms before making a buying decision.