Hark launches Handoff, a browser agent claiming speed and cost advantages over GPT-5.5 and Claude Opus

Hark, the startup behind a $700 million Series A round closed in May, is entering the crowded browser-automation market with Handoff, an agent designed to execute multi-step tasks on websites that lack public APIs. According to TechCrunch AI, the tool interprets visual layout and DOM structure to determine when to click buttons, enter text, or navigate pages — claiming both speed and cost advantages over established competitors.

Handoff’s technical approach and scope

Handoff operates by analyzing a website’s visual presentation alongside its underlying code structure, enabling it to identify and interact with interface elements without relying on documented application programming interfaces. According to TechCrunch AI, the system can work with major retailers (Target, Walmart), restaurant-reservation platforms (OpenTable), and professional networks (LinkedIn).

In a product demonstration shared by Brett Adcock, Hark’s CEO, the agent was shown completing a flower-arrangement order while interpreting imprecise user instructions—such as “some florist’s choice selections.” However, the preview covered only a portion of the full workflow, limiting assessment of how reliably Handoff handles multi-step sequences or error recovery.

Hark’s architecture diverges from conventional large language models in a fundamental way. Rather than predicting the next word in a sequence, Handoff predicts the next action—whether a mouse click at specific coordinates or a keyboard input in a particular form field. According to TechCrunch AI, this action-prediction model is paired with a post-trained foundation model, with plans to shift to pre-training later in 2026 to accelerate refinement of the data pipeline and training infrastructure.

The emerging agent infrastructure market

The browser-automation space is no longer a startup-only domain. Google, OpenAI, and Anthropic are all developing computer-use agents, joined by venture-backed teams including Browser Use, Polar, Strawberry, and Aside. Hark’s competitive positioning centers on two claims: superior speed relative to rivals and substantially lower operational cost compared to inference-heavy models like OpenAI’s GPT-5.5 and Anthropic’s Opus 4.8.

The startup has not released independent benchmarks validating these performance or pricing assertions. Hark has opened a public waitlist and plans to ship Handoff by the end of August 2026.

Why This Matters

Enterprise operations teams currently evaluating agent scaffolding for internal workflow automation—expense reporting, travel booking, vendor management—face a vendor-selection decision by Q4 2026 when pilot programs conclude. If Hark’s claimed cost advantage holds up under real-world deployment, it could shift the economic calculus away from token-per-second billing models toward action-prediction pricing. The distinction matters: an agent that performs task A in 10 actions costs far less than one requiring 100 token-generation steps to reach the same outcome. Watch whether Hark’s post-training approach yields faster convergence than pre-trained models from larger incumbents—that timeline and reliability record will determine whether the company captures share among teams building in-house agent infrastructure.