Hark shows off browser-use agent Handoff as it targets tasks websites still make hard

Hark has unveiled Handoff, a browser-based AI agent designed to complete online tasks on behalf of users, including booking tables, shopping, researching, and navigating sites that do not offer official APIs. The launch matters because Hark is betting that the next major frontier in AI is not chat, but action: software that can reliably operate websites the way a person would.

The preview comes just months after Hark raised a massive $700 million Series A in May, underscoring how much investor capital is still flowing into agentic AI. The startup says Handoff can interpret page structure and visual cues, then decide when to click, type, scroll, or fill in forms across services such as Target, Walmart, OpenTable, and LinkedIn.

That promise puts Hark directly into one of the most competitive corners of artificial intelligence, where a long list of well-funded companies and startups are trying to build agents that can do real work inside browsers rather than merely generate text. The company is also making a bold technical claim: its system predicts the next action, not just the next token.

What Hark is launching and why it matters

Hark is introducing Handoff as a browser-use agent meant to take commands and carry out multi-step web tasks without requiring users to do the clicking themselves. In practical terms, that means the system is aimed at workflows people already do online: placing orders, reserving restaurant tables, gathering information, and handling repetitive errands that often require manual navigation through websites.

The launch is important for two reasons. First, it highlights the shift from conversational AI to action-oriented automation. Second, it shows that Hark wants to compete on reliability and speed in a category where many demos look impressive but break down in real-world conditions.

Browser agents have become a popular pitch because the web remains fragmented. Many services still lack usable APIs, especially for consumer tasks. If an AI can operate a website the way a human does, it can potentially connect to thousands of services at once without waiting for each company to build an integration.

How Handoff works

Hark says Handoff reads both the structure of a website and the visual information on the page to understand what action to take next. That means the agent is not just parsing text in the abstract; it is trying to infer what a human would see and do, whether that is pressing a button, entering details into a field, or moving through a checkout flow.

The startup describes the model as a system that predicts the next action instead of the next word. That framing places Handoff in a different category from standard large language models, which are designed around token prediction. In Hark’s telling, the model is trained to decide on concrete interactions, such as selecting a clock time, clicking a form element, or typing at a specific location on the screen.

Why a browser-first approach is useful

A browser-first agent can, in theory, work across a wide range of websites without bespoke integrations. That is especially valuable for services like shopping, reservations, or travel, where each platform has its own interface and rules.

Hark argues that this approach makes the system more flexible than tools that depend on APIs. For consumers and businesses alike, the appeal is obvious: one assistant could potentially handle a range of tasks across multiple sites, even when no official developer support exists.

What the demo showed

In a video demonstration, Hark chief executive Brett Adcock showed the agent responding to a request to assemble a bouquet using user-specified flowers while also handling a vague instruction for “some of the florist’s choice.” The demo was meant to show how the assistant copes with ambiguity rather than only rigid instructions.

Still, the footage revealed only part of the workflow. Because the company did not show the full end-to-end process, it is difficult to judge how dependable the agent is in practice, how it handles errors, or how often a human may need to intervene.

Hark’s pitch is that the assistant can navigate ambiguity on the web and still carry out a useful task, but the company’s public demo does not yet prove how robust that ability is in real-world use.

That limitation is familiar in the agent market. Many systems work well in controlled demos but struggle with pop-ups, changing layouts, login prompts, payment steps, and sites that actively resist automated behavior. The real test for Handoff will be how it performs when users ask it to do ordinary work on messy, shifting websites.

Why Hark is using a post-trained model first

For this release, Hark says it is relying on a post-trained model rather than a fully pre-trained one. The company plans to move toward pre-training later this year.

That sequencing suggests Hark wants to refine its data collection, infrastructure, and training methods before investing more heavily in a larger base model. In the fast-moving AI market, this can be a practical strategy: ship a system, learn from usage, and then build the next version with cleaner data and improved techniques.

Hark says this approach gives it more room to improve the pipeline and training stack faster than if it tried to perfect everything before launch. For a company trying to establish itself in an increasingly crowded field, speed of iteration may be almost as important as raw model capability.

How Hark is positioning itself technically

The company’s core claim is that agentic systems should be evaluated by the actions they can take, not just the text they can generate. That sets up a philosophical and technical contrast with mainstream LLMs, which are typically measured by language outputs and benchmark tasks.

If Hark can reliably produce the correct next action in a browser session, it could reduce the need for hand-coded automation flows. But that also raises the bar for proof. Action prediction in the wild is difficult because websites are inconsistent, users are unpredictable, and small interface changes can break a workflow.

How does Handoff compare with other agent products?

Handoff enters a field that already includes offerings from Google, OpenAI, and Anthropic, along with startups such as Browser Use, Polar, Strawberry, and Aside. That competition matters because browser automation is quickly becoming one of the most visible race tracks in AI.

Most of these products share the same broad promise: give the model a goal and let it complete actions inside a browser. The differences usually come down to speed, cost, accuracy, how well they recover from mistakes, and whether they can handle websites that were not built with automation in mind.

Hark says Handoff is faster than rival systems and significantly less expensive than models such as GPT 5.5 and Opus 4.8. Those comparisons are likely to draw attention because price and latency are two of the biggest factors determining whether a task agent can be practical for everyday use.

What makes browser agents hard to build

Browser agents have to deal with the real internet, not a neat benchmark. That means dealing with logins, captchas, modal windows, slow pages, dynamic elements, and constantly changing site layouts.

They also need judgment. A good assistant must know when to keep going, when to ask the user for clarification, and when to stop because a task has reached an unsafe or ambiguous step. That is why many agent products still feel more like prototypes than full replacements for human effort.

• They must understand page layout and text.
• They need to react to visual changes in the interface.
• They must manage multi-step workflows without losing context.
• They have to fail gracefully when a site behaves unexpectedly.
• They need to do all of the above quickly enough to feel useful.

What kinds of tasks could Handoff perform?

Hark says Handoff is designed for a wide set of everyday web tasks. That includes shopping for essentials, ordering food or coffee, arranging travel, filing returns, reserving restaurant tables, and gathering research from multiple sources.

In principle, those are exactly the tasks that make browser automation attractive: they are repetitive, they involve structured forms and menus, and they often require more time than judgment. A reliable agent could save users a meaningful amount of effort if it can execute them without constant supervision.

Possible use cases

1. Buying items from retail sites that do not offer simple checkout APIs.
2. Booking reservations on restaurant platforms.
3. Collecting information across multiple websites for research.
4. Filling out forms for returns or routine administrative tasks.
5. Handling common consumer errands that require repeated navigation.

That said, the more valuable the task, the more important reliability becomes. A bot that can shop or book a table is only useful if it can complete the job consistently, securely, and without causing accidental purchases or incorrect reservations.

Why investors are paying attention

The scale of Hark’s May financing round suggests investors believe browser agents could become a major product category, not just a feature tucked inside chat tools. A $700 million Series A is exceptionally large by startup standards and signals an aggressive plan to build infrastructure, models, and product distribution around agentic automation.

Large funding rounds like that tend to reflect a belief that the market winner in AI agents could own a large share of future software workflows. If agents can reliably do the clicking, many existing services may become interfaces for machines as well as people.

That possibility helps explain why Hark is moving quickly to show product, even if the public demo is partial. In a competitive environment, visibility can matter almost as much as capability. Launching a named agent, opening a waitlist, and promising an end-of-summer release all help build momentum.

What comes next for Hark?

Hark says Handoff is now on a waitlist and is expected to ship by the end of the summer. That timeline suggests the company is still in an early access phase, using the preview to gather interest while it continues to refine the system.

The next few months will likely determine whether Hark can turn a promising demo into a product people actually trust. The company will need to prove that its agent is not only fast and cheap, but also accurate, durable, and safe enough for real consumer use.

The market for browser agents is becoming crowded, and Hark’s success will depend on whether Handoff can demonstrate reliable task completion outside of polished demos.

If it succeeds, Hark could become one of the better-known names in agentic AI. If it falls short, it will join a long list of tools that showed what browser automation might look like before the hard part began: making it work every day, for ordinary users, on the open web.

Timeline of Hark Handoff’s public rollout

The bigger picture for AI agents

Hark’s preview is part of a broader industry push to move AI from conversation into execution. The underlying idea is simple: if a model can understand a goal, it should be able to act on it. But the execution is much harder because the open web was not designed for autonomous software at scale.

That tension is likely to define the next phase of agent development. Companies will keep trying to build tools that can function like digital assistants, but the winners will be those that bridge the gap between impressive demos and dependable performance. Hark is now making its case that Handoff can be one of them.

For users, the promise is convenience. For the industry, it is a race to control how people and businesses will delegate work to AI. And for Hark, the preview of Handoff is a public bet that browsers may become the operating environment where the next generation of AI proves itself.