Hark previews its browser use agent for completing tasks

Seven hundred million dollars. Series A. Let that number settle. No AI startup in history has commanded a Series A at that scale — not OpenAI, not Anthropic, not the fleet of LLM companies that burned through billions before proving a single durable product. Hark's funding announcement in May should have triggered avalanches of skepticism. Instead, the tech press largely reproduced the figure. Today's launch of Handoff, a browser-use agent that supposedly navigates Target, Walmart, OpenTable, and LinkedIn without APIs, gives us the first chance to ask what that capital actually bought.

The premise is familiar. Issue a command — order coffee, book travel, file a return, research a topic — and the agent executes. Dozens of startups have pitched this exact vision. Browser Use, Polar, Strawberry, Aside. Google, OpenAI, and Anthropic have their own computer-use agents in various stages of release. The market is not waiting for Hark. The incumbents own the infrastructure, the talent, and the distribution. A newcomer needs either a fundamental technical breakthrough or a distribution hack. Hark offers neither in its launch materials.

Instead, the company leads with a demo video. CEO Brett Adcock instructs Handoff to build a bouquet with specific flowers plus "some of the florist's choice" — a fuzzy phrase meant to showcase natural-language flexibility. The video shows the agent navigating a florist site, clicking categories, selecting items. Then it cuts. We never see the checkout. We never see confirmation. We never see whether the order actually places. A demo that stops at the moment of truth is not a demo. It is a trailer.

Hark's technical claims deserve equal scrutiny. The company says it uses a "post-trained model" now and will pre-train later — framing the sequence as a strategic advantage that lets them refine data pipelines and training infrastructure faster. This inverts the standard paradigm. Pre-training builds the model's world knowledge. Post-training aligns it to specific tasks. Doing them backward suggests either a novel architecture or a fundraising narrative. The company also claims its model predicts "the next action — which could be a click or a keyboard input at a specific place" rather than the next token. That distinction matters if true. Action prediction could bypass the token-generation latency that plagues LLM-based agents. But Hark provides no benchmarks, no latency numbers, no comparison methodology. Only the assertion.

Then come the phantom competitors. Hark says Handoff is faster and "costs much less than other models like GPT 5.5 and Opus 4.8." Neither model exists. OpenAI has not announced GPT-5, let alone a 5.5 iteration. Anthropic's Claude lineup includes Opus as a codename for Claude 3, not a 4.8 version. Benchmarking against fictional products is a tell. It signals that the company either lacks real competitive data or hopes the press will repeat the claim without checking. So far, that hope looks justified.

The waitlist opens today. General availability targets end of summer. That timeline — months away for a product that just demonstrated a partial flower purchase — suggests the gap between demo and production remains wide. Browser automation is brittle. Sites change layouts daily. CAPTCHAs appear. Two-factor authentication interrupts. Session management fails. The agents that survive contact with the real web are the ones that have ingested millions of execution traces, not the ones that impressed in a controlled video.

Seven hundred million dollars buys compute. It buys talent. It buys time. It does not buy product-market fit. Hark now has the longest runway in the category. That runway lets them iterate in private while competitors ship in public. But it also insulates them from the discipline that comes from customers who can walk away. The waitlist is a Gateway drug — it measures interest without demanding commitment. Real traction comes when users trust an agent with their credit card, their calendar, their LinkedIn credentials. That trust is earned in production, not in a video that fades to black before the receipt appears.

The browser-agent category will consolidate. The winners will be the teams that solve the long tail of edge cases — expired sessions, dynamic selectors, rate limits, consent flows — not the teams with the largest war chests. Hark's capital is a fact. Its agent remains a claim. Until Handoff completes a task that matters, end-to-end, in front of strangers, the $700 million is just the most expensive seed round in history.