OpenAI launches Dots, always-on agents that keep working in the background
AI is moving from chat windows to agents that run for days with their own identities and tools, which makes permissions and memory the skills to learn.
OpenAI has launched Dots, always-on AI agents that pursue goals a user sets and keep working in the background with little supervision, according to reports by TechCrunch and The Register on 29 September 2026. The announcement came at the company's DevDay developer conference in San Francisco on Tuesday.
Dots run on GPT-6 Astra, OpenAI's latest model, and are available from Tuesday in ChatGPT and Codex for Pro and Business Premium subscribers in eligible markets. TechCrunch reports that OpenAI sees them handling jobs such as monitoring customer feedback, fixing bugs, or rerunning a scientific analysis when new data arrives. Users can name and customise a primary dot, and OpenAI pictures teams of specialised dots working at the same time. The agents connect to Slack, Teams and other workplace tools, with SMS support to follow, and OpenAI is working with Microsoft to integrate Agent 365 security controls. Individual dots can be given their own identities, credentials and tools through existing systems.
The Register reports that each dot gets its own cloud computer, can reach roughly 4,000 apps through connectors, and retains context over time as it learns a user's preferences. Chatting with a dot does not count against subscription usage limits, and OpenAI plans flat monthly fees for extra dots or faster ones. Enterprise customers can try a beta once an admin approves it.
OpenAI also announced GPT-6.1 Sol, which it says offers near-Astra intelligence at one-fifth of the price, a shared workspace called Space, and an Agents API that now supports computer use.
Why it matters
The chatbot that waits for a question is giving way to software that runs continuously on its own schedule. That shifts the hard problems from answering well to acting safely over hours or days: who the agent is, what it can touch, and how anyone checks its work. Giving each agent its own identity and credentials treats AI workers much like employees or service accounts. Learners who understand identity, permissions and long-running jobs will be well placed as this pattern spreads.
The chatbot that waits for a question is giving way to software that runs continuously on its own schedule.
Model: GPT-6 Astra
OpenAI's latest model does the reasoning for each dot.
Own cloud computer
The Register reports each dot gets its own hosted machine and keeps working toward its goal.
Memory over time
Dots retain context and learn a user's preferences and work patterns.
Connectors to about 4,000 apps
Integrations include Slack and Teams, with SMS support to follow.
Own identity and credentials
Dots can be given identities, credentials and tools, with Microsoft Agent 365 controls in progress.
Read from the top down. The widest layer is the base.
What you can learn from this
- Agent loops turn a model into a worker. A language model on its own takes text in and returns text out, once. An agent wraps it in a loop: the model reads the goal and the current state, chooses an action such as calling a tool, observes the result, and decides what to do next. Running that loop on a schedule, or when new data arrives, is what makes an agent always on. The model is the reasoning engine, but the loop, the tools and the stopping rules decide what the agent actually does.
- Memory is stored state, not a bigger brain. Models have a fixed context window, so anything an agent needs to remember across days has to be written somewhere and fetched back. Common approaches include saving notes or summaries, storing facts in a database, and retrieving relevant items into the prompt before each step. When OpenAI says dots retain context and learn preferences, the mechanism underneath is a system that decides what to save and what to recall. Poor choices here cause agents to forget key instructions or cling to stale ones.
- Connectors are how agents reach the world. An agent cannot do anything useful without tools, and connectors are the adapters that expose an app's actions, such as reading a channel or opening a ticket, in a form the model can call. Each connector usually sits on an API and an authorisation grant. Access to thousands of apps sounds like power, but every connection is also a set of permissions that can be misused if the agent is confused or tricked. That is why the scope of each connection matters as much as the number of them.
- Agent identity is becoming an access control problem. Giving each dot its own identity and credentials means actions can be traced to a specific agent instead of hiding behind a person's login. It also means each agent can be granted only what its job requires and switched off on its own. This is the same least privilege thinking used for service accounts in cloud systems, now applied to AI workers. The Microsoft Agent 365 integration points to agents being managed alongside human accounts in company directories.
- Cheaper models change what is practical. A model that runs around the clock consumes tokens constantly, so cost per token decides which tasks are worth automating. A model offering similar quality at a fifth of the price, as OpenAI claims for GPT-6.1 Sol, makes background monitoring jobs affordable that were not before. Many real systems route routine steps to a cheaper model and reserve the strongest one for hard decisions.
We teach this in the AI Tutor
How to use this in practice
- Map one task as an agent loop. Pick a recurring job you do, such as checking a shared inbox or a project board each morning. Write it out as four parts: the goal, the tools needed, the check that decides the next step, and the condition that means stop and ask a human. Done is a half-page description a developer could turn into an agent without asking you questions.
- Audit the app connections you have already granted. Open the connected apps or third-party access page in your Google, Microsoft or GitHub account settings. For each entry, note what it can do and when you last used it, and remove anything you do not recognise or no longer need. Done is a list that contains only apps you can explain, each with the narrowest access you can give it.
- Build a tiny scheduled agent with a hard boundary. Using any model API you already have access to, write a script that runs on a timer, reads one source such as an RSS feed or a local folder, and writes a short summary to a file. Give it no ability to send messages or delete anything. Done is a script that runs three times unattended and produces summaries you would trust, with its only permission being write access to one file.
- Compare costs before you automate. Take a task your script handles and estimate tokens per run by counting input and output characters and dividing by about four. Multiply by runs per day and by the published prices of two models you could use. Done is a small table showing the monthly cost of each option and which one you would choose.
Sources
- OpenAI launches Dots, its bubbly agentic avatarTechCrunch (opens in a new tab)
- OpenAI tries disarming AI angst with cute graphics and always-on agentsThe Register (opens in a new tab)
Our reporting is an original summary. Full coverage is at the links above.
How do you score in Agentic AI?
A few minutes of questions, then a report on where you stand and what to learn next.