The mascot is a distraction. OpenAI put a friendly blob on stage at DevDay 2026 and most of the coverage took the bait, because a cartoon is easier to write about than enterprise governance policy. But the actual story of Dots isn’t that OpenAI made agents adorable. It’s that OpenAI made agents that never stop running, gave them their own cloud computers, and then handed the keys to Microsoft’s IT admin console. That’s not a consumer product with a cute face. That’s infrastructure wearing a costume.
What was actually announced
On September 29, 2026, at DevDay in San Francisco, OpenAI introduced Dots: always-on agents that run on the GPT-6 Astra model, crawl the web continuously, and chase multi-step projects on a user’s behalf across apps. You can personalize them. They’re depicted as blobs. They’re rolling out in ChatGPT to Pro and Business Premium subscribers in eligible markets, with a beta for enterprise users.
The part that matters more than any of that: Dots plug into Microsoft’s enterprise governance and security controls through Agent 365, so businesses can manage them with the Microsoft tooling they already run. Reuters framed the launch as OpenAI going after Meta in the enterprise market, and that framing is closer to the truth than the mascot photos.
Why “always-on” is the whole review
Every agent I’ve tested over the past two years has had the same failure mode. You give it a task, it does something vaguely task-shaped, and then it stops. The stopping is actually a feature nobody admits to appreciating. When an agent halts, you get to inspect what it did. You get a checkpoint. You get a moment where a human decides whether to continue.
Remove the stop and you remove the checkpoint. An agent that constantly crawls the web on your behalf is an agent making decisions in the gaps between your attention. That’s the entire value proposition, and it’s also the entire risk surface, and those are not two separate things you can price independently.
So the questions I’d want answered before trusting one with anything that costs money:
- When a Dot does something wrong at 3 a.m., how do you find out? Does it tell you, or does it log quietly and keep going?
- What’s the rollback story on a multi-step project that went sideways on step four of nine?
- How does an always-on agent handle a web page that’s trying to manipulate it? Constant crawling means constant exposure to content that wants to redirect the agent’s goals.
- What does it cost when it’s running all the time? Idle compute isn’t free, and “always-on” is a billing model as much as an architecture.
OpenAI’s announcement says Dots get to know what matters to you. Fine. I want to know what happens when they get it wrong and nobody’s watching.
The Microsoft tie-in is the actual product
Agent 365 integration is the least glamorous line in the announcement and the most telling. Enterprises don’t buy agents because agents are impressive. They buy agents when the security team stops objecting. Wiring Dots into existing Microsoft governance controls is how you get past the security team. It means identity, permissions, and audit can be handled with tools the company already pays for and already understands.
That’s a sharp commercial move and a genuinely useful one. It also tells you who this is really for. Consumer Pro users get a blob they can name. Enterprise buyers get a managed agent fleet inside a compliance perimeter. Guess which one is the revenue story.
The honest verdict, with honest caveats
I haven’t tested Dots. It launched yesterday, it’s in staged rollout, and the enterprise version is beta. Anyone publishing a verdict on performance right now is guessing, and I’d rather tell you what I’m watching for than pretend I already know.
What I’ll be watching: whether continuous operation produces genuinely better outcomes than the task-then-stop model, or just more output. Those get confused constantly in agent demos. An agent that did forty things overnight looks productive right up until you check whether any of the forty were correct.
The optimistic read is that always-on solves the real problem with current agents, which is that they lose the thread the moment you close the tab. Persistent context and persistent execution are legitimately hard, and if GPT-6 Astra plus dedicated cloud computers cracks it, that’s meaningful.
The skeptical read is that OpenAI shipped a longer leash and a nicer mascot, and the hard parts of agent reliability are still hard. Both reads are currently supported by the same evidence, which is a keynote and a blog post.
Judge this one on the audit logs, not the blob.
🕒 Published: