Hark, a high-growth startup backed by $700 million in Series A funding, has officially previewed "Handoff," an intelligent agent designed to navigate the web just like a human. By moving beyond simple text generation, Handoff aims to automate complex workflows across websites that lack official API support.

One of the most significant hurdles in automation is the reliance on structured APIs. Hark’s Handoff agent bypasses this limitation by analyzing both the underlying website structure and visual data to interact with interfaces. This allows the agent to seamlessly navigate high-traffic, non-API platforms such as Target, Walmart, OpenTable, and LinkedIn.

The agent functions by understanding visual cues to determine when to click buttons, scroll, or input text. In a recent demonstration, CEO Brett Adcock showcased the agent’s ability to handle complex, "fuzzy" logic—such as building a custom flower bouquet where the user specifies "some of the florist’s choice." This capability suggests a level of semantic understanding that goes beyond rigid command execution, allowing for more natural human-to-agent interaction.

A Shift from Next-Token to Next-Action Prediction

Technically, Hark is pivoting away from the standard Large Language Model (LLM) paradigm. While traditional LLMs are optimized to predict the next token in a sequence of text, Hark’s model is architected to predict the "next action." This means the model specifically calculates the next physical or digital interaction, such as a specific keyboard input or a mouse click at a precise coordinate on the screen.

To accelerate development, Hark is currently utilizing a post-trained model. This strategy allows the company to rapidly refine its training infrastructure, data pipelines, and specialized techniques before moving to a full pre-training phase scheduled for later this year. This iterative approach is designed to optimize the agent for speed and cost-efficiency.

The Competitive Landscape of Computer-Use Agents

Hark enters a crowded and highly competitive arena. The race for "computer-use" dominance includes tech giants like Google, OpenAI, and Anthropic, alongside specialized VC-funded startups such as Browser Use, Polar, Strawberry, and Aside.

Hark is positioning itself as a disruptive alternative by claiming superior speed and significantly lower operational costs compared to heavyweight models like GPT 5.5 or Opus 4.8. By focusing on a specialized action-oriented architecture rather than a general-purpose text model, Hark seeks to capture the market for high-frequency, low-cost browser automation.

The company has opened a waitlist for its platform and intends to launch the full service by the end of the summer. For developers and founders, the success of Handoff could signal a shift toward "Action Models" that prioritize task completion over mere conversation.

Key Takeaways

  • Action-Oriented Architecture: Unlike standard LLMs that predict text tokens, Hark’s model predicts specific user actions like clicks and keystrokes.
  • API-Less Automation: The Handoff agent uses visual and structural data to navigate websites like Walmart and LinkedIn that do not offer official APIs.
  • Cost and Speed Focus: Hark aims to undercut the pricing and latency of major models like GPT 5.5, targeting efficient, large-scale browser task automation.

What to watch

  • Launch timeline – Hark has opened a waitlist and plans a full release by the end of summer.
  • Model evolution – Hark is moving from a post-trained model to a full pre-training phase scheduled for later this year.