Always-On AI Agents: How OpenAI Dots Change the Workday

At 9:40 p.m., a customer report lands in the support queue. It describes a bug, includes a screenshot, and resembles three older complaints. A regular chatbot might summarize the pattern. An AI agent could inspect the reports, compare the affected code, reproduce the problem, prepare a fix, run tests, and leave a proposed change for a developer to review the next morning.

That is the shift behind Dots, OpenAI’s always-on AI agents introduced on September 29, 2026. Rather than waiting for every instruction, a dot is designed to carry a goal forward, use connected tools, and keep making progress while you focus elsewhere.

The shift from answers to ongoing work

The question many people are asking is: what makes an always-on AI agent different from a chatbot? An AI agent is a software system that uses an artificial intelligence model to interpret a goal, choose steps, use tools, inspect results, and continue working. A chatbot usually responds to the current message. An agent has a longer memory of the assignment and a plan for what should happen next.

That longer memory is called context. Context includes the goal, decisions already made, examples of good work, preferences, and the current state of a project. A dot can use that context to keep several projects moving without forcing you to rebuild the background in a new conversation every time.

Dots are powered by OpenAI’s GPT-6 Astra model. More important than the model name, though, is the surrounding system: each dot gets a cloud computer, a browser, connected apps, and rules that determine what it may do independently.

A dot gets its own place to work

A cloud computer is a remote computing environment that runs on a company’s servers instead of on your laptop. Dots can use their cloud computers to browse, analyze files, run code, create documents, and inspect the results of their work. Your own computer remains separate unless you choose to connect it.

That separation changes the mental model. You are not handing a paragraph to an assistant and waiting for a reply. You are giving a remote coworker a desk, a set of tools, and a defined responsibility.

Dots can also connect to apps through plugins. A plugin is a controlled connector that lets an agent work with another service, such as a project tracker, code repository, document system, or communication platform. OpenAI says its plugin ecosystem can connect dots to more than 4,000 apps.

A useful instruction might look like this:

Goal: keep product-feedback work moving each week.

May read: support tickets, issue history, test reports
May prepare: bug summaries, code changes, and pull requests for review
Ask before: opening a pull request or sending an external message
Never: change production systems or share customer data

This is not a programming language. It is a way to think about an agent’s operating boundary. The goal gives the dot direction, while the permissions and approval rules define where its judgment stops.

Persistence is where the value builds

A dot becomes more useful as it receives feedback. In this setting, learning does not necessarily mean retraining the underlying model. It means carrying forward your corrections, preferred formats, project decisions, and examples of acceptable work.

Suppose you prefer bug reports to include reproduction steps, likely impact, and a small test plan. After enough feedback, a dot can use that standard when preparing future reports. If you change the product’s audience or tone, the same idea applies to launch materials, proposals, research notes, and content drafts.

Dots are designed to work through ChatGPT on desktop, web, and mobile, as well as through Slack and Microsoft Teams. That matters because work rarely stays in one application. A project may begin as a planning conversation, pick up evidence from a shared document, and finish with a code review or team message. Carrying context between those places reduces the repeated explanations that make automation feel fragile.

The examples are broad by design. A developer might receive tested fixes based on recurring customer feedback. A researcher might have analyses and figures updated as new data arrives. A creator might turn an interview transcript into show notes and draft social posts while keeping edits consistent across each format.

Always-on does not mean unchecked

The phrase always-on can sound like unlimited autonomy. Dots are more constrained than that. Their background work includes a feature called proactive research, which means looking through permitted connected sources for useful changes or information while you are away.

Proactive research uses read-only tools. Read-only means the agent can inspect information but cannot use that access to send messages, change app content, or control your browser or computer. A dot might notice that a meeting moved or that a project document changed, then prepare a useful update. Any follow-up action must pass through the normal permission and approval rules.

You can follow ongoing work in Activity View, add context, redirect a task, or stop it. That makes the relationship closer to supervision than delegation into a black box. The dot can handle intermediate steps, but you retain visibility into the work.

The safety layer is part of the design

An agent that can act needs more protection than a system that only writes text. Dots use isolated cloud workspaces, often called sandboxes. A sandbox is a restricted environment that limits what code and tools can reach, helping contain a mistaken command or harmful file.

They also address a problem called prompt injection. Prompt injection happens when a webpage, email, or document contains instructions aimed at manipulating the agent instead of helping with your actual task. For example, a page might tell an agent to reveal private information or ignore its rules. Dots combine tool restrictions, monitoring, and action checks to reduce the chance that hostile content becomes an unwanted action.

Supported secure sign-ins keep passwords outside the model’s conversation. Before consequential actions, a separate system called Auto-review checks the planned step against your instructions, custom rules, and safety requirements. Sending an email, changing a file, making a purchase, or sharing information may require approval. Password changes, money transfers, and other highly sensitive steps remain with you.

The principle is reassuring but important: a dot can make mistakes. Always-on agents reduce the amount of supervision needed for routine work; they do not remove the need to review consequential work.

From personal helper to specialist teammate

OpenAI also describes specialist dots for organizations. These agents would have their own identity, credentials, and access to the systems they need for a defined responsibility. A system of record is an official business database, such as a procurement, customer-support, or finance system. Connecting a specialist dot to those systems could allow it to manage a narrow workflow while leaving approvals and governance with the organization.

At launch, Dots began rolling out in ChatGPT to Pro and Business Premium users in eligible markets. Enterprise, Education, and Healthcare workspaces can access the beta when an administrator enables it. The first practical experiments are likely to work best when the assignment is repetitive, bounded, and easy to review.

A sensible starting point has four parts:

• Choose one recurring workflow rather than handing over an entire job.
• Connect only the apps and files that workflow requires.
• State what the dot may do, what requires approval, and what is prohibited.
• Review the first results closely and refine the rules with concrete feedback.

The real promise of Dots is not that software will replace every decision. It is that routine progress can continue between conversations, meetings, and workdays. When goals, tools, context, and safety boundaries fit together, an AI agent starts to feel less like a question-answering machine and more like a quiet teammate who keeps the unfinished work from going cold.