OpenAI's Dots agent aces own tasks, fails outside checks

OpenAI announced Dots earlier this week as an agent platform with blobby, anthropomorphic avatars and customizable names — closer in look to Meta's Muse than to conventional business software, but positioned squarely as work tooling. The Verge's Allison Johnson named hers Dotty McDotface and tested it on the $100-per-month Pro account she expensed, noting that Dots can access apps like Blender and GIMP right off the bat and can also be given access to a user's own computer through the desktop ChatGPT app.

The test exposed a split personality. When Johnson handed Dot tasks fully within her control — a personal website redesign and consolidating desktop video files — it delivered, even recreating a prototype block by block and pinging her on the ChatGPT phone app when the video clip was ready. Outside services were another matter: an internet installation booking stalled at a "human check" requiring a sustained mouse hold, an Ikea login looped in a security check, and the neighborhood teriyaki spot wouldn't let Dot onto its ordering page. A coworking space search also missed a free-tour option that Instinct found within minutes. Dot itself attributed the stalls to websites blocking its cloud browser traffic, and Johnson observed it required more manual takeovers than competing agents — possibly because it wasn't purpose-built for such errands.

The larger question is who Dots is for. The gated rollout to OpenAI's top-tier accounts, and the fact that users pay upfront rather than through eventual ad monetization, suggests an enterprise-first bet — a reading the product's mixed results on consumer errands only reinforce. Whether Dots becomes genuinely useful for a given worker may hinge on how much of their work lives inside systems they control, and on OpenAI closing the security-check gap that repeatedly forced Johnson to step in.