What changes when AI keeps working after you close the chat

What changes when AI keeps working after you close the chat?

For someone running a business or coordinating a team, the question is practical: which part of a task can move forward while you attend to something else, and what do you need to check when you return? Research, a comparison, or a draft can be good starting points. The usefulness of that delegation is assessed through the result and the work it took to review it.

On September 29, 2026, OpenAI introduced dots as agents that can keep work moving between conversations, with their own cloud computer and the apps the user connects. The announcement describes goals that continue to advance and results that come back for review.

Access is rolling out gradually. The official help distinguishes Pro availability by market, Business Premium, and an Enterprise beta enabled by an administrator. Before planning a trial, check availability in the account and the connections it requires.

One possible application for a business is to design delegation with continuity: define the goal, let work progress within its boundaries, and return to review what it produced. That possibility still needs testing in each team's actual work.

A request such as “help me with operations” leaves too many decisions open. Start with a task whose result you can recognize and review. For example, compare options for a specific need or prepare a summary from permitted documents.

Before delegating it, write down five things:

• The goal: which question you want to answer

• The sources: where it may look and what information it may use

• The deliverable: what it should return

• The limits: which actions it may take and which need approval

• The completion criterion: what it would need to show for the work to be considered finished

This also helps identify a poorly chosen task. If you cannot define how to review the result, reduce the scope before giving it more responsibility.

Imagine a team that wants to compare three tools for coordinating its weekly work. This is a fictional example to explain the method; it does not describe an Agentic Engineering implementation or a tested result with dots.

The assignment could be: “Compare three tools for a small team that needs to assign tasks and review progress. Use the providers' public documentation. Return a comparison with links, the limits of the plan you examined, and outstanding questions. Do not create accounts or contact the providers.”

The team would have a concrete review point: check whether the relevant features are documented, whether the conditions still apply, and whether the comparison addresses its need. If the focus drifts toward features that do not matter, the team can correct it. If information is missing, that absence should remain visible.

The expected result would be a comparison that helps the team decide. To know whether it achieved that, someone would need to read it, open its sources, and check it against the requirements. In this example, the choice of tool remains with the team.

Two situations should be distinguished. A task you have already assigned can continue within its authorization. OpenAI also describes proactive research that looks for useful information in permitted connected sources. The tools used for that research are read-only: they cannot send messages to other people, change app content, or control the browser or computer. Subsequent actions follow the usual rules and checks.

Closing the chat does not, by itself, expand the assignment. Authorization remains tied to what you requested. The official documentation also says that the rules continue to apply to delegated work and work that continues in the background. Consequential work requires review, and some sensitive actions need approval or must be completed by the user.

A simple review can start with three questions:

• Permissions and scope: did it have access to the necessary sources and stay within the assignment? If a connection failed or a document was left out, did it say so?

• Evidence: can you open the sources and check the important claims? Does it distinguish observations from an estimate or a proposal?

• Human decision: what still needs verification, and who decides the next step?

Ask the deliverable to make clear what was completed, what remains pending, and why. A list of activities does not prove that the goal was met. In the fictional comparison, finding three websites would be progress; checking the requirements and limitations would let the team assess the result.

In a first trial, also review how much correction work the team needed to do. That observation helps decide whether to repeat the assignment, adjust the instructions, or maintain closer review.

If the team has technical support, a bounded trial can record the inputs used, the available tools, the permissions granted, and the actions observed. Define acceptance criteria before starting and preserve enough evidence to check the result.

These are different states: having a tool available, attempting an action, receiving a response, and verifying that the expected change happened. Separating them helps locate errors and avoids treating a task as finished simply because the process kept running. This is a proposed evaluation criterion, not a description of dots' internal architecture.

Choose a low-risk weekly task and write down the result you would need to see to consider it useful. Start with research or preparing a draft, limit the sources, and define who will review it. After trying it, compare what you received with that criterion.

Which weekly task would you delegate to an AI agent, and what evidence would you need to trust the result? That is a useful question for starting to explore these tools with your team.