Remember when ChatGPT couldn’t remember you? Every session started cold. You re-explained your project, your preferences, your constraints, and then the window closed and the whole thing evaporated. We spent a good chunk of 2023 and 2024 building scaffolding around that amnesia — vector stores, retrieval layers, prompt templates that stuffed a user profile into every request. The model was a function call. Stateless in, stateless out.
On September 29, 2026, OpenAI announced Dots at DevDay in San Francisco, and the framing has inverted. Dots are always-on agents. They persist. They crawl the web on your behalf while you’re asleep. They’re depicted as cute blobs you can personalize and hand multi-step projects to. And per SiliconANGLE’s coverage, they come with their own cloud computers.
That last detail is the one I keep circling back to.
The architectural shift hiding behind the blob
An agent with its own compute environment is a different kind of system than an agent that runs inside your chat session. The distinction isn’t cosmetic. When a model executes inside a request-response cycle, its lifetime is bounded by yours. Failure modes are contained. If it loops, you see the loop. If it misreads a goal, you’re right there watching the tokens stream by.
Give that same model a persistent runtime and a mandate to chase goals across apps — Reuters described Dots as agents that “chase user goals across apps on their own” — and the bounding box disappears. The agent now operates on a timeline you’re not observing. Its errors compound in the dark. Its intermediate state lives somewhere you didn’t provision and can’t easily inspect.
This is the part of the always-on design that deserves the most scrutiny, and it’s the part the blob mascot is doing a lot of work to soften. A cute face on a long-running autonomous process is a UX decision with real consequences: it sets the user’s mental model to “pet” when the correct mental model is closer to “unattended cron job with a language model in the driver’s seat.”
Why Agent 365 matters more than the model
Dots run on GPT-6 Astra. Model capability is the headline, but the integration with Microsoft’s enterprise governance and security controls in Agent 365 is the more interesting engineering signal. Businesses can manage Dots through existing Microsoft tools.
Think about what that requires. To govern an always-on agent, you need answers to questions that stateless chat never forced anyone to formalize:
- What identity does the agent act under, and is it distinct from the human who created it?
- What is the scope of its credentials, and does that scope shrink when the task completes?
- Where does its activity get logged, and at what granularity — task level, tool call level, token level?
- Who can pause it, and what happens to in-flight work when they do?
- How do you revoke an agent that has been running for three weeks and has touched a dozen systems?
These are identity and access management problems, not machine learning problems. The fact that OpenAI shipped Dots with enterprise governance hooks rather than bolting them on later suggests the company understands where the actual adoption blocker sits. Enterprises don’t hesitate on agents because the models aren’t good enough. They hesitate because nobody can tell them what the audit trail looks like.
Personalization as a load-bearing component
OpenAI’s announcement describes Dots as getting to know what matters to you. In a stateless system, personalization is a nice-to-have — a better starting prompt. In an always-on system, it becomes load-bearing. The agent is making decisions without you in the loop, which means its model of your preferences is the only thing standing between “helpful” and “confidently wrong for six hours.”
That raises a question I’d want answered before deploying these at scale: how does a Dot represent uncertainty about your intent, and what does it do when that uncertainty is high? A well-designed always-on agent should have a clear escalation path — a way to stop and ask rather than guess. Whether Dots have that, and how aggressive the threshold is, will determine whether they’re useful or exhausting.
What to watch during the rollout
Dots are rolling out in ChatGPT to Pro and Business Premium users in eligible markets, with a beta for enterprise users. Staged access like that is the right call for a system class this new, and it gives the rest of us a window to observe before committing architecture decisions around it.
The thing I’d measure first isn’t task success rate. It’s failure visibility: when a Dot gets something wrong, how quickly does the user find out, and how much work does it take to understand why. Agents that persist are agents whose mistakes persist. The interface can be as friendly as it likes, but the debugging story is what determines whether this generation of always-on systems earns a place in production or stays a demo.
đź•’ Published: