Cute Blobs With Cloud Computers Are Now Your Coworkers

Almost every bot I’ve shipped has been a vending machine. You put in a prompt, something falls out, the machine goes back to sleep. Dots are not vending machines. Dots are more like a housecat with credentials — awake when you aren’t, wandering the house, occasionally knocking something off a shelf on your behalf.

OpenAI announced them at DevDay 2026 on September 29 in San Francisco, and the framing is deliberate: always-on agents, drawn as cute blobs, that you personalize and hand multi-step projects to. They run on GPT-6 Astra, they crawl the web continuously to push tasks forward, and per SiliconANGLE they come with their own cloud computers. Reuters describes them chasing user goals across apps on their own, which is the part that should make anyone who has built an agent loop sit up straight.

Always-on is an architecture, not a feature

If you’ve built agents, you know the shape of the standard loop: user message in, plan, tool calls, results, response, done. The session ends and so does the agent’s existence. Every hard problem in that design is bounded by the length of one conversation.

Remove the boundary and the problem set changes completely. An agent that never stops needs answers to questions a chat turn never asks:

• When does it wake up? Polling, event triggers, or its own judgment about what’s worth checking again. Each choice has a different cost curve.
• What does it remember? Long-lived state means stale state. An agent working a project for a week accumulates assumptions that quietly stop being true.
• What stops it? A loop with no user turn to interrupt it needs budget ceilings, step limits, and a definition of “done” that isn’t “the human replied.”
• What does it touch? Crawling the web on your behalf means outbound requests you didn’t personally authorize, reading content you didn’t personally vet.

That last one is the part I keep circling. An always-on agent reading the open web is an always-on agent ingesting untrusted text. Any builder who has watched a scraped page try to talk their agent into something knows this isn’t theoretical. The safe posture is to treat every crawled document as data, never as instruction — and to assume that boundary will be tested constantly rather than occasionally.

The Microsoft integration is the tell

Dots plug into Microsoft’s enterprise governance and security controls in Agent 365, so businesses can manage them with tools they already run. That detail reads like plumbing, and it’s the most informative fact in the whole announcement.

It tells you OpenAI understands the actual blocker for persistent agents in a company. Nobody’s IT director is worried about whether the blob is cute. They’re worried about an unattended process with access to internal systems, no audit trail, and no off switch that a compliance team recognizes. Answering that with existing Microsoft management tooling instead of a bespoke console is a pragmatic call. It means the agent shows up in the same place as every other managed identity.

For those of us building outside that stack, it’s also a spec sheet by implication. If you ship an always-on agent, someone will eventually ask who can create one, what it can reach, what it did at 3 a.m., and how to revoke it. Building those answers in after the fact is miserable. Build them first.

Cute blobs are a serious design choice

Mashable’s coverage includes a yellow blob mascot named Alfred, and it’s easy to read the whole aesthetic as marketing gloss. I think it’s doing real interface work. A persistent process is hard to reason about. A character is not. Giving an always-on agent a face and a name gives users a mental model for something that otherwise has no natural representation in a chat window — it’s present, it’s yours, it’s doing something.

Personality also sets expectations about fallibility. You forgive a pet for knocking something over. You don’t forgive a cron job. That’s a useful buffer when the agent’s job involves autonomous judgment across a week of work.

Where this leaves builders

Dots are rolling out in ChatGPT to Pro and Business Premium users in eligible markets, with a beta for enterprise. WIRED frames them as OpenAI’s answer to Meta’s Muse, which means the always-on agent category now has two large players competing on it rather than one experimenting with it.

My read: the interesting work for independent builders isn’t cloning Dots. It’s the supporting layer. Persistent agents need scheduling that doesn’t burn money, memory that expires on purpose, permission scoping that survives an audit, and observability that lets a human reconstruct a week of autonomous decisions. Those are unglamorous, solvable, and mostly unbuilt outside the big labs.

Start there. The blob will handle the cute part.

🕒 Published: September 29, 2026