SmartAGI · News · 1 October 2026
On 29 September in San Francisco, OpenAI introduced Dots: agents the company describes as “always-on”, able to keep tasks moving while the user turns to something else. DevDay also brought separate announcements: an Agents API that can operate software, third-party extensions for ChatGPT, automations triggered by events in other apps, and shared spaces for teams. The ambition is clearer when the announcements are taken together: OpenAI wants work to begin in ChatGPT, continue there while the user is away, and return when there is something to decide.
“Always-on” is the promise, not universal availability. Dots are offered to Pro and Business Premium customers in eligible markets. Enterprise, Education and Healthcare workspaces can try the beta if an administrator enables it; it is off by default. That distribution matters. The same assistant may be available to one user and not another, while an organisation can decide which features are enabled and which connections are allowed. Automation is no longer just a model capability: it depends on the account, subscription and rules set by whoever administers the workspace.
The clearest step from answering to acting comes with computer use in the Agents API. A developer can build an agent that interacts with software to complete tasks. OpenAI lists tool calling and search, coordination among multiple agents and context compaction. In partnership with Amazon, it is also offering managed agents on AWS Bedrock. The chat window becomes the entry point to sequences that cross applications and services.
Answers and actions carry different consequences. A wrong answer can lead someone to make a bad decision; a connected agent can execute steps inside an account, repository or business workflow. For Dots, OpenAI says each agent works on its own cloud computer, separate from the user’s unless the user connects them; background research uses read-only tools; and people can choose apps, set rules to allow or block actions or require approval, and follow progress in Activity View. The company says its auto-review checks actions that could affect accounts or share information, while some sensitive tasks, such as changing a password, remain with the user. These are controls described by OpenAI, not an independent assessment of how well they work. Its product page explains part of Dots’ permission system; the recap does not specify what complete action history is retained for each task or how every operation can be undone. When an agent works while we are not watching, that record is no interface extra: it is how we find out what happened in our account.
The company is bringing the same logic to professional work. Codex Security Cloud can scan repositories, investigate findings, group duplicates and prepare fixes, even with a laptop closed. Code Review lets users inspect changes and ask Codex for a first cloud-based review. Decisions API, still in limited preview at the end of the event, routes requests or chooses an agent’s next step from answers defined by the developer. These are separate products with separate availability: Code Review is listed for all plans; Security Cloud for Pro, Business, Enterprise and Education; a wider release of Decisions API was promised for the following days.
The model also reaches into organisations’ daily work. ChatGPT Space brings a team, its documents and an agent together; members can share pages and knowledge, and ask the system to organise the space according to their instructions. On mobile, users could find, read and share, while creation and editing were still to come. Pages adds collaborative documents; shared slides were still due in the coming weeks. Based on a still-proposed specification, MCP Events would let automations start when something happens in a connected app, such as a new task appearing on a project board. “Available”, “beta” and “coming soon” are not footnotes: they define what a user can do today and what the company is promising.
OpenAI also wants third parties inside the same surface. Plugin extensions let developers add interactive panels and a place in the sidebar. Users choose which plugin to open and approve the access it receives. Sign in with ChatGPT lets people use their account identity to sign in to outside services; for Plus and Pro users of participating tools, it can also make part of their plan allowance available. OpenAI says 16 partners are involved. Its enterprise Marketplace does something else: eligible customers can put part of their existing OpenAI commitments towards approved software from an initial group of 32 partners.
That design could save integrations, passwords and steps between apps. It also turns distribution into a source of power. A provider that controls the assistant, identity, recommended plugins and place where agents run can influence which tools customers find and the conditions under which they connect. OpenAI limits the Marketplace to approved partners and eligible enterprise customers; it also lets them put part of existing commercial commitments towards those products. The assistant is not only recommending software: it sits between the buyer and the contract that pays for it.
ChatGPT wants to become the environment that keeps context, calls software, assigns tasks and manages connections. When an agent acts for a person or a business, people need to reconstruct what it did, limit in advance what it could touch and correct the result. Dots includes access controls and an activity view, according to OpenAI; their reach will have to be tested in practice. “Always-on” describes the agent’s schedule. Its activity record will have to describe its conduct.
Sources: DevDay 2026 Recap, OpenAI · Introducing Dots, OpenAI · Official DevDay programme