OpenAI's always-on agents: what dots change for a small team this month

The interesting part of the dots announcement is not the avatar. It is the persistence model: OpenAI describes dots as always-on agents built to handle ongoing work and keep making progress between conversations, rather than waiting for your next prompt.

What actually shipped

From OpenAI's own release notes: dots are powered by GPT-6 Astra, each has its own cloud computer, and you create one in the ChatGPT desktop app or on desktop web, connect apps, and define what it can do on its own.

The availability is the part to read twice. Rollout is to eligible Pro and Business Premium users in supported markets, with Pro access excluding the EEA, Switzerland and the UK at launch. Enterprise access is a beta and off by default. So if you are on a business plan and waiting for it to appear, it may simply not be on for your account yet.

What it changes if you run a small team

One sentence: the cost of automation moves from writing prompts to supervising something that does not stop.

That is genuinely different for an owner-run shop. Most small teams abandon AI workflows not because the model is weak but because someone has to sit down, reopen the chat, paste the context and restart the job. If you or your office manager spend about ten minutes a day doing that, five days a week, persistence is roughly an hour a week back. Run that arithmetic on your own minutes, not mine.

What it does not change: whether the output is right. I have not tested dots, and OpenAI's page does not claim the agent will be correct unattended.

The permission ladder to write before you connect anything

This is the checklist I would use, and it takes about thirty minutes to fill in:

1. Scope. Name the one job. Sorting an inbox, drafting a weekly summary, chasing a status. Nothing touching money or customers in week one.
2. Read only first. Connect with read access before you allow a single write action.
3. One app. One connected app, not your whole stack. Blast radius stays small.
4. Owner. One named person reads the agent's output every morning for the first week.
5. Stop switch. Write down who can turn it off and how long that takes. If nobody knows, you do not have one.
6. Failure log. A place to record what it got wrong, dated. Two weeks of that tells you whether to widen the scope.

If you cannot fill in lines four and five today, do not connect anything yet.

What unattended actually breaks

When I built a closed-CRM bridge across 25 GoHighLevel locations, the failure that cost the most was not dramatic. It was 201 responses that drop data: the call looks successful, the record is not there, and an unattended process carries on as if everything is fine.

An always-on agent has the same shape. It will not stop to tell you it is confused. Whatever you give it, assume it will keep doing for a week before anyone checks. More on how I think about staged permissions on the blog.

Limitations

• I have not run a dot. Everything above about behaviour comes from OpenAI's pages, not from my own test.
• OpenAI's exact limits for heavy usage, add-on pricing and long-term commercial terms were not confirmed in the primary sources I could read, so budget cautiously past the first month.
• Governance is not shipped for you. Who authorises an action, who monitors output, who handles a mistake: those are your rules to write.
• Community reaction has been loud, but reaction is sentiment, not evidence. Nobody has run these in a small business for a quarter yet.

My take

This is the right direction and the wrong moment to bet a process on it. Persistence is the missing piece for teams with no staff to operate software, and my guess is that within a year the useful agent is defined by its permissions, not its prompt.

So: one dot, one job, read only, one owner, one stop switch. Review it in two weeks. If it is boring and correct, widen it.