OpenAI just released Dots, and it introduced a completely different way to work with AI...
Dots is in a category of its own right now, not because it chats better than ChatGPT or codes better than Codex, but because it does not try to replace either
out of the box it’s an always-on agent with its own cloud computer, access to more than 4,000 apps, and the ability to keep working across your browser, files and tools even when you’re not there
and if you understand how to give it responsibilities, context, permissions and approval rules, you can hand over parts of your work that were previously too ongoing, fragmented or unpredictable to automate
people are calling it the first real AI that works for you even when you’re not there for a reason...
so this is the full course
what Dots is actually exceptional at, how to use it beside ChatGPT and Codex, the responsibility patterns that pull the best work out of it, the permission gates that stop bad actions, and the five workflows where it can make or save real money
one note: Dots is still rolling out gradually, Pro access currently excludes the European Economic Area, Switzerland and the UK, and Enterprise access requires an admin-enabled beta
this is early software, so start with reversible work and review anything consequential before you let it touch something expensive
but if you want to understand the agent everyone is suddenly talking about, start here
First: do not use your dot like a normal ChatGPT chat
A normal conversation is useful when you want an answer right now. A dot makes more sense when the work continues after the answer.
https://x.com/OpenAI/status/2104984507107717331
You can ask ChatGPT to summarise one customer interview. You can give your dot responsibility for watching new feedback, finding repeated requests, scoping smaller fixes, and bringing you tested pull requests with videos showing what changed.
You can ask ChatGPT to rewrite one launch email. You can teach your dot the audience, positioning and creative standards for the entire launch, then let it update the materials when the product changes.
You can ask ChatGPT what is on your calendar tomorrow. You can ask your dot to review it every morning, notice conflicts, prepare what you need and remind you when something deserves attention.
The difference is not the size of the prompt. It is the type of responsibility.
If the work is...Use your dot to...ExampleOngoingKeep track of the goal as new information arrivesWatch customer feedback and surface recurring problemsRepetitiveRun the same useful process on a scheduleReview tomorrow's calendar every eveningSpread across toolsCoordinate connected apps, files, ChatGPT Work or CodexTurn each new transcript into clips, notes and social draftsChanging over timeUpdate the work instead of starting againRevise a proposal when customer requirements changeDependent on your judgmentPrepare the work and stop at the decisionResearch options, draft a recommendation and ask before committing
The first rule is simple: stop giving your dot isolated prompts and start giving it a job.
1. GIVE IT A RESPONSIBILITY, NOT A TO-DO LIST
Most people will start by sending their dot a giant list.
Research these competitors. Read these emails. Update this document. Check my calendar. Draft five posts. Remind me on Friday.
It may complete the list, but you have not really created an agent. You have created a longer chat request.
A better starting point is one area of responsibility. Tell the dot what outcome it should protect, what information matters, how you judge the work, and which decisions still belong to you.
OpenAI's examples all follow this pattern. A developer's dot watches feedback and prepares tested fixes. A scientist's dot updates an analysis when new evidence arrives. A sales lead's dot keeps a proposal current as the customer's requirements change. A creator's dot turns every new transcript into clips, notes and social drafts.
None of those are one-off questions. They are standing jobs.
The Dots playbook
Use this structure:
Responsibility:
[the part of the work your dot owns]
Goal:
[the result it should keep moving toward]
Context:
[the people, product, project and previous decisions it should understand]
Standard:
[what good work looks like]
Autonomy:
[what it may research, organise, draft or change on its own]
Approval rule:
[what it must show you before acting]
Update rhythm:
[when it should check, report or ask]
Here is a strong first responsibility:
You are responsible for keeping our launch materials accurate as the product changes. Use our existing docs, approved copy and relevant Slack discussions as context. You may find inconsistencies and prepare updated drafts on your own. Do not publish, message creators or change live pages without my approval. Check for meaningful changes every weekday and only alert me when something needs a decision.
The trick is not clever wording. It is giving the dot a real outcome, the facts it cannot guess and a line it cannot cross.
Start narrower than you think. "Help me run my company" is not a responsibility. Customer feedback, one launch, one content pipeline, one recurring report or one active sales process can be.
You can widen the job once the dot has shown that it understands the work. It is much easier to expand a reliable system than to debug an agent that was given half your life on day one.
2. TEACH IT WHAT GOOD LOOKS LIKE
Dots learn from feedback over time, but only if your feedback contains useful information.
"This is bad" teaches very little. "This sounds too polished and generic; keep the concrete example, remove the dramatic opening, and write it like I am explaining the idea to another founder" teaches a reusable preference.
The same applies to research, code, design and analysis. If you reject a proposal, say whether the problem was the evidence, risk, price, timing or recommendation. If you rewrite an email, explain why your version works better. If you choose one design, point to the hierarchy, spacing or visual tone you prefer.
The dot is trying to learn what done the way you would do it means. Your job is to make that standard observable.
Give it examples, not adjectives
Words such as clean, premium, natural and strategic mean almost nothing by themselves.
Instead of saying:
Make our launch copy sound natural.
Say:
Use these three approved posts as the voice reference. They are direct, slightly informal and explain the product through a concrete problem. Avoid corporate language, fake excitement and claims we cannot prove. When uncertain, prefer the shorter sentence.
Instead of saying:
Bring me good leads.
Say:
A good lead is a Shopify brand with strong repeat-purchase behaviour, visible mobile traffic, active product drops and no existing mobile app. Exclude stores that appear inactive or sell products people rarely reorder. Return the evidence for every qualification.
Examples give the dot something it can compare against. Adjectives force it to guess.
When the same mistake appears twice, stop fixing the individual result and update the standing instruction. Each correction should make the next piece of work better, not only repair the current one.
3. CONNECT THE MINIMUM IT NEEDS
Dots can work across connected apps, their own browser, their cloud computer and, if you explicitly allow it, your local computer. That access is what makes them useful, but it is also what makes setup matter.
Do not connect everything just because the option exists.
Start from the responsibility and work backward. If the job is turning interview transcripts into content, the dot may need the folder where transcripts arrive, your approved examples and the place where drafts belong. It probably does not need your personal email, calendar and every company workspace.
If the job is keeping a proposal current, it may need the account history, product documentation, relevant messages and proposal files. It does not automatically need permission to send emails or change the CRM.
The rule is simple: connect the smallest useful set of information and tools, then expand when a real task requires it.
Connection and permission are not the same thing
Your dot might be able to read a connected inbox without having standing permission to send messages. It might prepare a document without being allowed to share it. It might research a purchase while still needing your approval before buying.
You manage connected accounts through ChatGPT's plugin controls. Disconnecting an app stops new information from being shared through that connection, but it does not remove information the dot has already learned. To erase the dot's own conversations, saved memories and scheduled tasks, the current reset flow deletes the dot.
That detail matters. Disconnecting is not the same thing as forgetting.
Every dot has its own protected cloud computer. Access to your local computer is optional and starts turned off. The cloud computer is enough for browsing, analysing information, creating files and running tools. Connect your own computer only when the job genuinely requires local files, your local browser, installed tools or local skills.
4. USE SCHEDULED WORK AND PROACTIVE RESEARCH DIFFERENTLY
One of the most interesting things about Dots is that the agent can help without waiting for a new message.
There are two different ways this happens.
Scheduled work
Scheduled work is something you deliberately ask the dot to run at a particular time or repeatedly.
Examples:
- Review tomorrow's calendar every evening and flag anything that needs preparation.
- Check the weekly support summary every Monday and identify the three fastest-growing complaints.
- Reconcile two revenue dashboards on the first day of each month.
- Review the active launch checklist every weekday until launch.
You can manage active, paused and completed schedules from the dot's profile, including their timing and notifications.
A strong scheduled task has a reliable trigger, a stable process and a clear output. "Think about my business every morning" has none of them. "Every weekday at 9am, review yesterday's new support conversations and show me only product issues mentioned by at least three customers" has all three.
Proactive research
Proactive research is different. When you are not actively working with your dot, it can look through permitted connected information for ways to help.
OpenAI says this background research uses restricted read-only tools. It cannot directly send messages, change connected app content, or control a browser or computer. It gathers information and keeps private notes for the dot. Any later action still has to follow the usual permissions and safety checks.
Treat proactive research as an early-warning system, not invisible automation.
Use it to notice a drifting deadline, a repeated customer problem, new evidence that changes an analysis, a forgotten invoice, or a calendar change that affects a trip. Then let the dot bring you the observation, prepare the next step and follow the action rules you set.
5. DESIGN THE APPROVAL RULES BEFORE SOMETHING GOES WRONG
The best way to improve your dot is not to make it autonomous everywhere. It is to decide precisely where autonomy helps and where it creates unnecessary risk.
Custom Rules give you four behaviours:
- Take action without asking
- Take action if pre-approved
- Ask before taking action
- Hand off to you
"Pre-approved" means you explicitly asked for the action in your prompt.
Use independent action for low-risk, reversible work: organising research, preparing drafts, rerunning an analysis or updating a private checklist.
Ask for approval when the action affects another person, spends money, publishes information, alters shared work or is difficult to undo. That includes sending an email, sharing a file, publishing a post, making a purchase, changing a live page or deleting data.
OpenAI says purchases made using saved cards require approval. Some actions, such as permanent deletion, running software from an unrecognised source or granting new security-sensitive access, require confirmation each time. Other steps, including changing a password or transferring money between financial accounts, must be handed back to you.
A good default is:
Research, analyse, organise, draft and prepare freely. Ask before sending, spending, publishing, sharing, deleting or changing access.
Dots also use a separate system called Auto-review to check certain planned actions against your instructions, Custom Rules and built-in safety requirements. It can allow an action, block it or make the dot ask you for more information.
That is useful protection, but it does not replace clear instructions. OpenAI explicitly warns that dots can still make mistakes, so consequential work still deserves review.
THE FIVE RULES THAT MAKE DOTS DRAMATICALLY BETTER
- Assign a continuing outcome
"Write a launch plan" creates a document. "Keep our launch plan accurate as the product changes" creates a responsibility.
- Teach through examples and decisions
Give the dot approved work, rejected work and the reason behind your choices. Every correction should improve the next result.
- Connect access gradually
Start with the smallest set of apps and information needed for the job. Keep local computer access off until a task genuinely requires it.
- Make actions and approvals explicit
Say what the dot may do independently, what needs prompt-level pre-approval, what it must ask about and what always stays with you.
- Review at decision points
Do not micromanage every click. Ask the dot to return when the recommendation changes, it needs new access, another person would be affected, or a complete result is ready.
That is how you get autonomy without losing control.
A SIMPLE WAY TO CHOOSE YOUR FIRST DOT RESPONSIBILITY
Pick something that is:
- repeated often enough that the dot can keep helping;
- clear enough that you can recognise good work;
- connected to information the dot can access;
- reversible enough that an early mistake is manageable; and
- annoying enough that removing it gives you real time back.
Good first responsibilities:
- Keep a launch plan and its materials aligned as the product changes.
- Review new customer feedback and prepare evidence-backed product recommendations.
- Turn every new interview transcript into clips, notes and social drafts.
- Review tomorrow's calendar and prepare what each important meeting needs.
- Keep one active sales proposal aligned with the latest customer requirements.
Bad first responsibilities:
- "Run my company."
- "Handle all my messages."
- "Buy whatever I need."
- "Make every decision I do not have time for."
- "Connect to everything and find ways to help."
Those prompts hide the goal, provide no standard and give the dot no reliable way to know where human judgment matters.
For your first run, finish the responsibility brief with one extra instruction:
Show me your plan before taking any action.
You will catch missing access, vague standards and wrong assumptions while they are still cheap to fix.
THE POINT IS NOT TO HAVE AN AI THAT IS ALWAYS BUSY
It is to have an agent that keeps the right things moving when your attention is somewhere else.
Dots are OpenAI's answer to a different kind of AI relationship. Your dot learns the goals and standards you teach it, works across its own computer and the apps you connect, keeps progressing between conversations, and returns when it has useful work or needs your judgment.
The people who get the most from Dots will not be the people who connect the most apps or write the longest prompts. They will be the people who define a responsibility clearly, show what good looks like, and design the right boundary between autonomous work and human decisions.
Start with one responsibility. Give it the minimum access it needs. Keep consequential actions behind approval. Review the result, explain your decisions and improve the standing instructions.
Then let the second run be better than the first.
That is how a dot stops feeling like another chatbot and starts becoming genuinely useful.
Availability at launch
Dots began rolling out on September 29, 2026. OpenAI says they are gradually becoming available to Pro users outside the European Economic Area, Switzerland and the UK, to Business Premium users across supported ChatGPT regions, and to Enterprise users through an admin-enabled beta. Access may take several days to appear.
You create your first dot through the ChatGPT desktop app or desktop web. At launch, it can then be used through supported mobile experiences, but a dot cannot be created on mobile and dots are not supported on mobile web. Your first dot is included with eligible Pro or Business Premium access at no extra cost.
