Browser automation is suddenly crowded, but Hark believes it has found an edge: predicting actions rather than tokens. The startup, which banked $700 million in Series A funding back in May, unveiled Hark Handoff on Wednesday, an agent designed to navigate websites that lack official APIs. Its target list reads like a daily shopping routine — Target, Walmart, OpenTable, LinkedIn. Handoff studies a page's structure and visual cues to decide whether it should click a button or type information. That approach lets it handle commands that would normally require human judgment, like ordering coffee, booking travel, filing returns, or researching across multiple sources. Main Developments CEO Brett Adcock demonstrated the agent in a video, tasking it with building a bouquet from flowers the user specifies. The system handled fuzzy language, interpreting phrases like "some of the florist's choice" without needing exact instructions. That flexibility is a step beyond rigid command parsing, though the demo only showed part of the process, leaving its full effectiveness unverified. Read also: Zoox Robotaxis Enter Paid Service: What Changes in Vegas Hark is releasing Handoff with a post-trained model, a deliberate choice that diverges from the pre-training route many rivals take. The company argues this lets it refine its data pipeline, training infrastructure, and techniques faster, iterating on real-world performance before scaling up. Full pre-training is slated for later this year. The startup also claims a fundamental architectural difference from typical large language models. While LLMs predict the next token, Hark's model predicts the next action — a click or keyboard input at a specific location. That design, the company says, makes Handoff both faster and significantly cheaper to run than competitors like GPT-5.5 and Opus 4.8. Background Hark isn't entering an empty field. Google, OpenAI, and Anthropic all have computer-use agents in development, each betting that browser-based task automation will be a key interface for AI. VC-backed startups like Browser Use, Polar, Strawberry, and Aside are also chasing the same opportunity, creating a crowded market where differentiation is hard to prove. The $700 million Series A in May positioned Hark as a serious player before it even shipped a product. That funding gave the company runway to develop Handoff quietly and now launch it with a waitlist rather than a general release. The timing also matters — the browser agent space has matured enough that early demos no longer impress on novelty alone. Why It Matters If Hark's action-prediction approach holds up, it could undercut the economics of AI task automation. Cost and speed are the two barriers keeping browser agents out of everyday use, and the company is directly targeting both. Cheaper agents would make routine automation practical for individuals, not just enterprises with big AI budgets. The focus on sites without APIs is also strategic. Most consumer-facing platforms resist building public interfaces for automation, so agents that can navigate them visually unlock far more real-world utility than those limited to well-documented services. That's the difference between a demo and a tool people actually use. What's Next Hark has opened a waitlist for its platform, with a planned release by the end of the summer. That timeline gives the company a narrow window to convert interest into a working product before rivals ship their own improvements. The post-training phase will likely reveal how well the model handles edge cases the demo didn't cover. Pre-training later this year remains the bigger question. If the company can scale its action-prediction approach with the same efficiency shown in the post-trained version, it could reset expectations for what browser agents cost to run. Until then, the waitlist will test whether demand matches the hype.