August 5, 2026
Hark, the robotics and AI startup led by CEO Brett Adcock, unveiled a browser-based AI agent called Handoff, according to TechCrunch. It's built to complete real tasks on ordinary websites -- ordering food, booking travel, filing returns, making restaurant reservations -- by reading a page's structure and visuals to decide where to click or type, without needing an official API from sites like Target, Walmart, OpenTable, or LinkedIn. Hark's pitch is speed and cost: the company claims Handoff is faster and “costs much less” to run than models like GPT-5.5 or Opus 4.8, though TechCrunch notes it hasn't published benchmarks to back that up. Hark currently runs a post-trained model and plans to pre-train a purpose-built one later this year -- notably, it says the model predicts the next action (a click, a keystroke) rather than the next token, a different bet than how most coding and browsing agents are built today. Hark raised $700 million in a Series A in May 2026; Handoff is on a public waitlist now, with a full release targeted for later this summer.
Spotify is building a paid add-on that lets listeners generate AI covers and remixes of real artists' music, and Merlin -- the licensing collective representing more than 30,000 independent labels and distributors -- has joined Universal Music Group in backing it, according to TechCrunch. What's notable isn't the feature itself, it's the guardrails Spotify built in before shipping: artists have to opt in, they get credited, and they get paid out of the tool's revenue -- Spotify doesn't even need a full catalog to launch, just the artists who agree. Co-CEO Alex Norström summed up the approach as wanting “real artists, not fake artists” at the center of it. That's a direct answer to the wave of AI-music tools built on scraped catalogs with no consent and no payout, and it's the same consent-plus-compensation model any AI product touching someone else's intellectual property should be judged against.
Put side by side, Hark and Spotify are making opposite bets on what makes an AI product trustworthy. Hark is racing on raw capability -- ship an agent that can act on the open web faster and cheaper than the competition, and prove the guardrails work later. Spotify did the opposite: it built the compensation and consent model first, then went to labels for the content. Neither approach is wrong on its own, but they point at the two places any AI product lives or dies right now -- whether it actually works reliably at the task, and whether the humans whose data, content, or accounts it touches were ever asked. Worth asking of any AI tool a business adopts: which of these two paths did the vendor take, and which guardrail did they build in before shipping rather than after?
Sources