OpenAI used DevDay on September 29 to announce the thing everyone sort of expected and nobody quite priced in: dots, always-on agents that live inside ChatGPT and keep working between your conversations. Not a chat mode. Not a scheduled task. An agent with a job description that doesn’t sleep.
What a dot actually is
Each dot runs on GPT-6 Astra, OpenAI’s flagship, and gets four things no ChatGPT feature had before:
- Its own cloud computer. A real machine it controls, not a sandbox that evaporates when you close the tab.
- Its own browser. It reads the web like you do, on its own schedule.
- Context that travels. The same dot reaches you in Slack, Teams, and ChatGPT and remembers what it promised yesterday.
- A permission layer. You decide what it does without asking, and everything else waits for approval.
That last bullet is the one that matters. The difference between a demo and a coworker is what the agent is allowed to do unsupervised. OpenAI built the guardrails into the product instead of leaving them as homework.
Why this changes the math
I wrote a few weeks ago that we should stop ranking models and start ranking cost per finished job. Dots are the logical endpoint of that argument. You don’t prompt a dot. You hand it an outcome, and it comes back with work. The billing question stops being tokens and starts being salary, and that is a much healthier conversation for buyers.
The early reviews are doing what early reviews do. Some people call it incredible. Some people can’t figure out what to delegate. Both are right. A dot is only as good as the responsibility you’re willing to hand over, and most of us have never practiced handing anything over.
The catch
Always-on means always-spending. An agent that checks your accounts every few hours burns tokens whether or not anything happened. Set the permission layer tight, give each dot one clear job, and audit what it did this week before you trust it with next week.
My take
Dots are the first product from a big lab that treats the assistant as staff instead of software. Meta has Muse coming, Anthropic is circling the same space, and xAI is building its own version. The agent war didn’t start this fall, but this is the season everyone showed up with uniforms.
If you run a small team, start one dot this week on something boring and measurable. A weekly competitive scan. An inbox triage pass. Judge it like you’d judge an intern: by what it finished, not how it sounded.