A product brief changes, and three drafts need another look. An ordinary answer can describe the mismatch. An ongoing assignment gives an agent a reason to return when the source changes again. OpenAI introduced Dots on September 29 for work across conversations, with a separate cloud computer and connected apps. OpenAI says it runs on GPT-6 Astra. The Verge's launch report places the announcement at DevDay 2026.

The useful distinction is what happens between requests. A dot can follow changing information, prepare a revision and return with a result or a decision. It is an AI agent , pursuing a goal through actions. OpenAI’s examples describe intended capabilities and early use, rather than independently measured reliability across every task.

The computer, apps and messaging channel are separate connections

Each dot has a cloud computer, browser and tools. With the access you grant, it can work with connected material, create files and analyze data. You can inspect its activity or redirect the work. Connecting your own computer is a separate, optional permission. OpenAI’s setup guidance distinguishes those connections from the channels, including ChatGPT, Slack and Teams, through which you can talk to the dot.

OpenAI calls its background reading proactive research. It can examine permitted connected sources and keep private notes without a fresh request. Those research tools cannot send messages, change app content or control a browser or computer. A follow-up action goes through the applicable instructions and checks. That separates learning about a change from acting on it.

OpenAI’s two-minute launch video shows a website update, a board-meeting deck, an app migration and personal scheduling. The launch film gives those continuing jobs a visible shape, moving from new information toward another piece of work.

OpenAI introduces Dots and shows example projects. Watch on YouTube

Six ways to give the continuing work a shape

The following proposed assignments give that capability a concrete shape. Each has a continuing goal, named inputs and a result someone can inspect. OpenAI gives examples across development, research, launches, sales and content work. The exact assignment still needs to fit the material you connect.

Turn customer feedback into a tested fix. Give the dot access to a feedback channel and a software project. Ask it to collect repeated complaints, reproduce one small defect, make a branch, run the relevant tests, and bring back a pull request with a short demonstration. The useful output is a reviewable change and a record of what it tested. A developer still decides whether the diagnosis is right and whether the change should be merged.

The comparison with Claude Duet takes the next step by asking which parts of a job can run while you are away and which decisions should wait.

Keep an analysis current. Point the dot to a defined dataset, an existing analysis, and the research question. Ask it to check for new data, rerun the same method when data changes, investigate surprising differences, and prepare revised figures with a change log. The result is more useful than a fresh summary because it shows what moved and why. A researcher checks the data, method, and interpretation before using the figures in a paper or decision.

Follow a launch through scope changes. Connect the approved product brief, release plan, and draft communications. Ask the dot to compare each new product decision with the current FAQ, launch email, and landing-page copy. It can prepare proposed edits and flag claims that no longer match the product. The team approves the final promise to customers and any publication or outbound message.

Process a recurring content intake. When a new interview transcript arrives, ask the dot to identify moments worth clipping, supply timestamps, draft show notes, and prepare a few posts in an established voice. The editor gets a package to review rather than starting every asset from a blank page. Check quotations and context against the recording, then approve any public post.

Catch an administrative loose end. An independent professional could ask a dot to compare completed assignments with an invoicing log, prepare a missing invoice from approved terms, and flag a due date. OpenAI describes an early tester whose dot prepared an overlooked invoice and sent it only after approval. That approval boundary is important: finding an omission and drafting the document are different from sending a financial request to a client.

Coordinate a personal plan when details change. A household could give a dot the approved schedule and a short list of vendors, then ask it to find a replacement if a booking falls through, compare available times, and propose a revised plan. OpenAI's launch film depicts a canceled wedding cake vendor and a replacement tasting. In a real assignment, the person should approve any new booking, payment, or message before the dot commits to it.

These proposed workflows share a pattern: something changes, the agent has relevant context, and a reviewable result follows. A one-off factual question or short rewrite may already fit an ordinary ChatGPT conversation. The ongoing job is the reason to consider something that works between them.

A small first assignment

Try one changing source and one output. For example: “Watch the approved launch brief until Friday. If a requirement changes, draft the corresponding FAQ and email edits. Show the differences and ask before publishing or messaging anyone.” Check the first result against the changed requirement: did it find the affected promise, update the right passages and leave unrelated copy alone? The source, duration, output and action boundary give that review something specific to inspect.

Current setup and the work allowance

The current OpenAI documentation supports creation in the ChatGPT mobile app, desktop app or a desktop browser. Mobile setup needs the latest app. Mobile web is not supported. Access is rolling out gradually. Eligible Pro users aged over 18 outside the European Economic Area, Switzerland and the UK can receive access. Business Premium is rolling out across supported regions, and an Enterprise administrator must enable dots for the workspace.

Conversations with a dot do not count toward ordinary ChatGPT usage limits. Work it starts or manages in Codex or ChatGPT Work uses those products’ normal limits. Plans include an allowance for deeper work, with extended limits for the first month after launch. Check that allowance against the amount of ongoing work you intend to assign.

Keep the connections as specific as the job. A messaging channel, an app plugin and a personal computer each provide different access. Use Activity to inspect work, and Custom Rules for a recurring action boundary when one is useful. OpenAI’s safety explanation identifies prompt injection in outside material as a risk, alongside remaining mistakes. The recipient, file change or final draft is still worth checking against the requested result.

For the launch-brief example, the first useful result is small: one changed requirement, the corresponding draft edits and a clear account of what is ready for a decision. That gives the next conversation somewhere concrete to begin.

FAQ

What is a reasonable first assignment for a dot?

Pick one source to follow, a result you can inspect and a clear stopping point before it sends or publishes anything. Choose connected apps appropriate to that task.