OpenAI Dots Working While You Play Solitaire?
OpenAI proposes agents that keep making progress between conversations. The practical question is what they can complete while you are away—and what still needs your judgment.
Yes, according to OpenAI: Dots are designed to keep making progress between conversations using a cloud computing environment. The idea is to leave a task running while you do something else—even play solitaire. The agent still depends on online infrastructure, authorized access and the conditions required to perform the work. Here, “offline” refers to the user's absence, not an agent operating without internet access.
Introduced on September 29, 2026, Dots are agents powered by GPT‑6 Astra, with their own cloud computer and access to the applications users choose to connect. The proposal is to pursue goals over time rather than limit interaction to a question and an answer.
The more interesting question comes afterward: when we return, do we find useful work completed, a draft requiring review, or a task waiting for a decision? Our experience with other AI tools highlights the importance of defining requests clearly and reviewing deliveries. That experience is not a test of Dots.
What Does Working While We Are Offline Mean?
It helps to separate human presence from technical execution. A person can leave the browser or stop following a conversation and return later. A cloud agent can continue an authorized task without waiting for that person to write the next message.
That does not mean every task will finish without intervention. An incomplete request, an unavailable connection or a decision requiring approval can prevent progress.
Continuous availability does not guarantee uninterrupted execution. It does not mean unlimited capacity either. The practical question is whether the agent has enough information, tools, authorization and resources to perform and complete the requested work.
An objective such as “help me with the website” is broad and leaves many decisions open. A request such as “review these three articles, identify broken links and prepare a report without publishing changes” provides a verifiable outcome and a clear boundary. Being direct and supplying the necessary context helps an agent understand what it should deliver.
What Are Dots?
OpenAI presents Dots as persistent agents that learn preferences through work and feedback. The launch includes interaction through ChatGPT, Slack and Teams, with a gradual rollout across eligible plans and markets.
Continuity is central to the proposal. Returning to a project with tasks organized, questions identified and materials prepared can be valuable. That progress is one of the main benefits we expect from delegation.
Knowing preferences does not mean perfectly understanding every intention. An editorial style, for example, involves choices that can change from article to article. An agent needs to distinguish previous habits from current instructions.
To assess usefulness, look at the outcome: how much work was completed correctly, and how much review time was required?
Continuing a Task and Researching Proactively Are Different
The documentation distinguishes user-requested work from proactive research. In the latter mode, background tasks consult connected, authorized sources through read-only tools. They save notes for the agent but cannot directly message other people, modify applications or control a browser or computer.
Connecting an application therefore does not automatically turn every piece of information encountered into an instruction to act.
For readers, there are two separate questions: “Can the agent consult this information?” and “Can the agent make this change?” A positive answer to the first does not settle the second.
How Could This Help a Small Website?
Consider an illustrative example, not a test conducted by NTS. A website owner asks an agent to compare a draft against supplied sources and prepare a list of claims requiring confirmation.
While the owner handles another activity, the goal would be to advance that review and leave a result for later inspection. The task is clearly bounded: verify and prepare, with publication reserved for a subsequent decision.
A second example would be analyzing reader requests and organizing topic suggestions. The useful outcome would not necessarily be choosing the next article independently. It could be presenting options with clear reasons and enough evidence for the editor to decide.
These examples illustrate a practical approach to delegation: hand over time-consuming stages while preserving the points where editorial judgment matters. Actual execution depends on access, tools, limits and product availability.
Who Decides What an Agent Can Do?
According to OpenAI, users choose connected applications and can define Custom Rules. These rules can guide actions, require approval or block behaviors within the product's mandatory protections.
The company also describes separate checks before certain actions and mandatory confirmations for operations such as purchases and permanent data deletion. Some sensitive steps, including changing passwords or transferring money between financial accounts, remain with the person.
In practice, authorization should specify the desired outcome and relevant boundaries. “Prepare a reply” and “send a reply” are different requests. It also matters who may receive information and what can be shared.
For a business, that distinction helps prevent a good intention from becoming a decision nobody intended to delegate.
Working in the Cloud Does Not Automatically Provide Access to Your Computer
OpenAI describes a separate environment for each Dot. Your personal computer remains separate unless you choose to connect it. That connection is an additional option, subject to applicable permissions and safeguards.
We should therefore not assume that a task involving local tools will continue when the required device is switched off. Execution depends on where the necessary resources are located.
Working without the user's presence is different from working without access to a required file, application or device.
What Are the Risks of Leaving Work Running?
OpenAI acknowledges that Dots can make mistakes. It also describes protections against malicious instructions encountered in webpages, emails and documents, known as prompt injection.
A generic example explains the issue: a document consulted by an agent could contain a sentence attempting to order the disclosure of private information. That text should be treated as source content without gaining authority over the user's request.
There is also an oversight challenge. The more steps an agent performs, the more important it becomes to understand what was done, which sources were used and where questions arose.
Three questions help when reviewing a delivery: does the result address the request? Are the claims supported? Did the agent remain within the defined boundaries? A persuasive answer does not replace these checks.
“24/7” Does Not Mean Unlimited Work
The announcement describes continuous availability while distinguishing it from the resources allocated to deeper work. Access depends on the plan, market and, within organizations, administrative settings. Limits and conditions should be checked at the time of use.
To evaluate value for money, the relevant measure is the cost of each accepted outcome. A quick delivery requiring extensive corrections may save less time than it appears to.
Being able to delegate many tasks does not mean you must. Starting with a clear objective and assessing quality can help identify where continuity offers a real advantage.
NTS View
In our assessment, the Dots proposal changes how work is organized: users can leave an interaction and return to useful progress. The value lies in continuity between those moments.
Autonomy becomes useful when it is connected to clear goals, verifiable outcomes and well-defined decision responsibilities. Leaving a task running requires understanding what has been delegated and what still depends on us.
For small projects, the opportunity could involve preparing research, reviews and materials while the owner is busy. Success should be judged by the quality of the work found on returning, including the effort needed to approve it.
NTS has not conducted an independent evaluation of Dots for this article. This analysis relies on manufacturer documentation and illustrative usage examples.
In our team's reported experience with other AI tools, there are already situations where we leave a task running and return to organized materials aligned with the request. That reinforces the importance of clear instructions, defined objectives and human review. Dots propose extending this continuity, but direct evaluation is needed to determine what the product adds to work we can already delegate.
Yes, the proposal allows work to continue while the user is offline. The decisive question is whether that work remains correct, authorized and useful when the user returns.