AI glossary · Basic · Agents
Always-on agent
An always-on agent is an AI agent that does not only respond each time the user talks to it. It stays running in the cloud, remembers the user's goals, moves work forward on its own, and reaches out first when it has a result or needs a decision.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Perception & action
- Fields
- Agents
- Added
- 2026-10-16
What it is / how it differs
A typical chatbot acts only when asked a question, and the work stops when you close the chat window. An always-on agent runs on its own computer and browser in the cloud, so it can keep working even when the user's computer is off. It can handle several projects at once and sometimes finds and handles necessary tasks before the user asks. A representative example is dots, released by OpenAI on September 29, 2026, which OpenAI introduced as an 'always-on agent'.
How it differs from a typical chatbot
Typical chatbotAnswers when asked
- Acts only when asked a question
- Stops when the chat window closes
- A person instructs each step
Always-on agentStays on and keeps working
- Finds needed work first
- Continues in the cloud even when my computer is off
- Reaches out when results or decisions are needed
공통 Both are built on LLMs, and a person makes the final judgment
How it works
First, the user connects apps such as email, messengers and document tools to set the scope the agent can use. The agent looks through the connected material and first finds and organizes useful changes, which dots calls 'proactive research'. It then carries out the work and contacts the user when an action or deliverable needs human approval. The user's feedback and conversation context accumulate in memory, and the more they work together, the better the agent understands the user's preferences and the standard for good work.
The work loop of an always-on agent
User's goal
- 1Observe first
- 2Do the work
- 3Request approval or decision
- 4Apply feedback
- 1. Observe first Find changes in connected apps
- 2. Do the work On its own computer and browser
- 3. Request approval or decision Reach out first when needed
- 4. Apply feedback Remember preferences and standards
Why it matters
Until now, using AI has mostly meant a person giving instructions one step at a time and checking the results. An always-on agent carries several steps forward on its own once given a goal and criteria, so people can focus on review and decisions. OpenAI said dots started investigating as soon as a bug was posted in Slack, and noticed an invoice an early tester had forgotten, drafted it, and sent it after getting approval. Cases like these show AI shifting from a 'tool' to a 'coworker'.
Things to know (limitations)
Because it is always on and accesses many apps, permission management is central. dots says it sets a default for each action—ask for approval every time, approve in advance, or act without asking—and that users can add rules such as "Never send emails." Its proactive research tools are blocked from sending messages to other people, changing content in apps, or operating the browser or computer. It also says memory is shared with ChatGPT, and that turning off sharing does not delete information that has already been shared. As of October 2026, dots is being rolled out gradually to Pro, Business Premium and Enterprise plans in some regions, and people under 18 cannot use it.
Example
OpenAI's dots runs on GPT-6 Astra, and whether you talk to it in ChatGPT, Slack or Teams, the same agent answers with the same memory. OpenAI said it can connect to more than 4,000 apps through its plugin ecosystem. For an app developer, for example, a dot finds recurring requests in customer feedback, implements and tests a small fix, then brings a PR with a video showing the changes and asks for review. Meanwhile, the developer can keep working on the next major feature.