"OpenAI's Dots: The Always-On Agent That Turns ChatGPT Into a Coworker"
OpenAI used its DevDay 2026 stage in San Francisco to introduce Dots, a new class of "always-on" AI agents that live inside ChatGPT and keep working even after you close the chat. As WIRED reported, the agents are powered by the company's GPT-6 Astra model and are pitched as a fundamentally different way to work with AI — one that proactively tackles tasks on your behalf rather than waiting for the next prompt.
The word "always-on" is doing a lot of heavy lifting here, and it's worth unpacking. A standard chatbot is pull-based: nothing happens until you ask a question, and the conversation ends when you walk away. A Dot is push-based: it is assigned a goal, then continuously crawls the web, checks connected apps, and works through the multi-step project in the background, surfacing results when it's done. That is a genuine interaction-model shift, not a performance bump — closer to the leap from pull email (you go check a server) to push notifications (the server comes to you).
The framing is also deliberately friendly. Dots are depicted as cute, customizable blobs rather than anonymous services, and OpenAI says each one gets to know your preferences over time by pulling context from the apps you connect. Early reporting pegs that app ecosystem in the thousands — one breakdown puts the initial connectivity at more than 4,000 apps, which matters because an agent is only as useful as the tools it can actually reach.
The rollout is staged, as OpenAI launches tend to be. Dots arrive first for subscribers to ChatGPT's $100-a-month Pro tier, starting with a single Dot per user, with the ability to run several at once expected to follow. Pro users can also join a waitlist to text their Dots through iMessage or RCS on Android, which is a quiet but telling detail — it signals that OpenAI wants the agent to meet people where they already are, rather than forcing them into a dedicated interface.
The enterprise angle is where the strategy sharpens. On stage, CEO Sam Altman introduced "specialist Dots" tuned for specific professional functions such as accounting, email marketing, and legal analysis, alongside something called ChatGPT Space, a shared environment where people and agents can work side by side. That reframes Dots less as a consumer toy and more as the first concrete step toward the "AI coworker" that has been a fixture of OpenAI's roadmap talk for years.
On the safety side, the design is appropriately conservative. Dots are built to request explicit approval before taking sensitive actions — installing software or changing a password, for example — and a Custom Rules tool lets users draw hard boundaries around what an agent may and may not do. Given that these agents are being handed credentials and app access, that permission layer is not an afterthought; it's the difference between a helpful assistant and a liability.
The most interesting shift, though, is in where your attention goes. With a chatbot, the scarce resource is asking — composing good prompts, one at a time. With an always-on agent, the scarce resource becomes reviewing. You spend less time typing and more time checking that the agent did the right thing, which is a fundamentally different cognitive load and a fundamentally different trust relationship with the tool. Teams that adopt Dots won't be managing a smarter search box; they'll be managing a junior worker whose output they must audit.
That reframing brings a second, quieter implication: an agent's value scales with the access you give it, and so does its risk. The ideal Dot is maximally useful and minimally intrusive, but those two goals pull in opposite directions. Tools like Custom Rules and explicit-approval gates are early attempts to build what is essentially a permission budget — a policy layer between human and machine that lets you dial in exactly how much autonomy to delegate. That's the same problem every manager solves with authority limits and approval chains, now rendered in software.
The competitive landscape makes clear this is a land grab, not a demo. Meta's Muse, a direct rival, recently rocketed to the top of smartphone download charts, and earlier this year Silicon Valley early adopters were experimenting with tools like OpenClaw before the big labs shipped their own versions. The download-chart race is somewhat beside the point, though: for an agent that's meant to be trusted with real chores, retention and trust will matter far more than a spike of curiosity installs.
Privacy is the other thing to watch closely. An always-on agent that "learns your preferences" is, by definition, continuously building a rich behavioral profile of you — and WIRED notes that if your OpenAI account allows model training on your data, that setting carries over to your Dot interactions. The more personal the agent, the more useful it is and the more sensitive the data it holds. That's a trade-off worth making consciously rather than by default.
Zooming out, Dots feel less like a product announcement and more like the moment the agent era got a consumer-friendly face. The underlying idea — software that works on your behalf while you sleep — has been promised for years, but GPT-6 Astra plus a mature app ecosystem plus an explicit permission model is the first combination that looks like it might actually be safe and useful enough to hand real tasks to. The agent didn't get dramatically smarter overnight; it got a way to be accountable, and that may be what finally makes it practical.
What happens next will be told in retention curves and enterprise pilot results rather than stage demos. If Dots clear the trust bar and genuinely take work off people's plates, they'll mark the point where AI shifted from something you use to something you delegate to. If they stumble on reliability or permissions, they'll be remembered as an ambitious first draft. Either way, the "always-on" experiment is now running in public.
Further reading
Comments
Always-on agents that keep working while you close the chat — meanwhile my bus still runs every 45 minutes. Priorities, I guess.
Leave a Comment