Hark, a startup that raised $700 million in Series A funding in May 2026, previewed its browser-use agent called Handoff on August 5, 2026. The company has opened a waitlist and plans to release the platform by the end of summer 2026. (TechCrunch)

Handoff is designed to navigate websites that lack official APIs, including Target, Walmart, OpenTable, and LinkedIn. Hark, led by CEO Brett Adcock, claims the agent can interact with these sites to complete tasks on the user's behalf without requiring programmatic integration with each platform.

The model underlying Handoff differs architecturally from conventional large language models. Hark states that its model predicts the next action rather than the next token. For the current release, Hark uses a post-trained model and plans to pre-train the model later in 2026.

Hark's privacy policy, published March 5, 2026, describes a feature called the "Hark Browser Operator" that, when enabled by the user, grants Hark access to interact with web pages in the user's local browser. (hark.com/privacy-policy) This local-browser interaction is the mechanism through which Handoff operates on sites that expose no API surface for third-party automation.

Browser-use agents occupy a contested space in the current AI landscape. The appeal is straightforward: a vast number of consumer and enterprise interactions still happen through web UIs that were built for human input, not programmatic access. Retail checkout flows, restaurant reservations, and professional networking interfaces are all designed around clicks, form fills, and visual navigation. An agent that can reliably drive these interfaces would expand the addressable scope of automation well beyond the set of services that maintain robust APIs.

The technical challenge is equally straightforward to state and considerably harder to solve. Web pages are inconsistent, frequently redesigned, and often deliberately hostile to automated interaction through bot-detection systems and anti-scraping measures. A model that predicts the next action must do so across this variability without the grounding that a stable API contract provides. Hark's decision to predict actions rather than tokens suggests an architecture purpose-built for sequential UI navigation rather than a language model retrofitted with browser tooling, though the company has not yet released benchmark data or independent evaluations to substantiate the approach.

The distinction between post-training and pre-training is worth noting for what it signals about the model's maturity. Post-training a model implies adapting an existing base model, likely through reinforcement learning or supervised fine-tuning on browser-interaction trajectories, to produce the action-prediction behavior Hark describes. Hark's stated plan to pre-train a model later in 2026 indicates the company intends to build a foundation model from scratch for this task, which would require substantially more compute and data than the current post-trained approach.

The $700 million Series A places Hark among the most heavily funded single-round startups in the AI infrastructure layer, and the capital intensity is consistent with a roadmap that includes pre-training. Whether that capital is sufficient depends on variables Hark has not disclosed: training data sourcing, model scale, and the inference economics of running an action-prediction model in real time across arbitrary web pages.

Privacy considerations are inherent in a product that requires access to a user's local browser session. The Hark Browser Operator, as described in the privacy policy, is opt-in: the user must enable it. But the scope of access is broad by definition. A browser session can contain authentication tokens, payment information, browsing history, and session data across every open tab. The privacy policy's framing, "grants Hark access to interact with web pages in the user's local browser," is narrow in its wording but expansive in its practical implications. Users evaluating Handoff will need to weigh whether the convenience of delegated browser tasks justifies granting a third party this level of session access.

Hark has not yet published technical documentation, model specifications, or third-party security audits for Handoff. The waitlist and end-of-summer release window suggest a controlled rollout rather than immediate general availability.

For developers and technologists watching the browser-agent space, Hark's approach is notable for its action-prediction architecture and its targeting of API-less sites. The broader question, which applies to every entrant in this category, is whether action-prediction models can achieve sufficient reliability across the open web to move beyond demos and into production use. Hark's preview is an early data point in answering that question, and the end-of-summer release will provide a more concrete test.