OpenAI Dots: The Agent Department Is Now a Line Item

OpenAI launched Dots at DevDay yesterday, and most of the coverage is about the avatar. I read the launch as an org chart change, because that is what it is for anyone running a product team.

The short version

Dots are always-on agents on GPT-6 Astra, each with its own cloud computer and browser, connected to more than 4,000 apps, included free with ChatGPT Pro and Business Premium, with specialist dots for companies in enterprise pilots and coming to Microsoft Agent 365. The agent department, the second unit in the two-department model I have argued since the pod died, just became a line item on a plan any employee can switch on. Three things product owns by Friday: a named owner per dot (13 of my 39 agents are orphaned, and a named owner was the best predictor of survival), one org-wide stop rule instead of a per-user setting (money, contracts, individual customers, and compounding decisions get a human gate a user cannot turn off), and a designer on the rules surface, because the approval prompt and the wrong path are the interface. For enterprise buyers, a specialist dot is an FDE deliverable sold direct, and the promise still needs the one-page handshake. The number nobody launched with is how many dots still do production work at ninety days.

What shipped

From The Next Web and TechCrunch, both read in full. Each dot has its own isolated cloud computer and browser. It connects to more than 4,000 apps through plugins, works on several projects at once, and can be messaged or called in ChatGPT, Slack, and Teams, with texting coming. When you are not working with it, it does what OpenAI calls proactive research, using connected apps in read-only mode, so it cannot send or change anything on its own. One early tester's dot noticed an unsent freelance invoice, pulled the details from the email thread, drafted it, and sent the PDF after approval.

Rules ship with it. Built-in defaults on when to act alone and when to ask, custom rules per user to allow, block, or require approval for specific actions, an auto-review step for anything that affects accounts or shares information, and a few tasks, like changing a password, that always stay with the user. A monitoring system can pause a dot. And OpenAI's own sentence: "Dots can still make mistakes, so always review consequential work."

The first dot is free with Pro and Business Premium. Pro users in the EEA, Switzerland, and the UK do not get it for now. Enterprise, Edu, and Healthcare workspaces get a beta once an admin turns it on. Specialist dots, with their own identity, credentials, and access to company systems, were tested internally on procurement, invoice processing, email marketing, customer support, and commercial contracting, and are starting as enterprise pilots and through Microsoft's Agent 365. OpenAI's line on where this goes: "Over time, we envision teams of dots working together on your behalf."

TechCrunch's read is that most of this was already possible through Codex and similar harnesses, and Dots bundles it with a focus on independent action and a friendlier face. That is fair, and it is exactly why the org question arrives now. Capability that needed a builder is now a toggle.

The agent department is a SKU

In The Triad Is Dead. Pods Are Dead. Agent Departments Now. the frame is two units: a human department doing judgment work and an agent department doing routine and synthesis work, coordinating on exceptions through a published interface. Humans on the legacy clock, agents on a continuous clock.

Yesterday the continuous clock became a plan feature. A dot runs while its owner sleeps, reads the company's connected apps, and drafts. Teams of dots are the stated roadmap. The question a product leader gets to skip is "should we build agents." The question they cannot skip is who owns the ones that exist on Friday, because by Friday they will.

Rule one: a name on every dot

39 PM AI Agents Deployed: What Stuck, What Died, and Why is my own fleet, 39 agents in 80 days. Thirteen are orphaned, with nothing anywhere referencing them. The single biggest predictor of survival was whether a named human owned the agent. The biggest predictor of death was scope drift, an agent that started focused and ended up doing five jobs badly.

A dot is built for scope drift. It learns your preferences, looks for ways to help in the background, and takes on more projects as you hand them over. That is the feature. Without an owner it is the failure mode. So the first rule is boring and non-negotiable: every dot in the workspace has a named human whose name is on its output, and a dot with no owner gets turned off, not adopted.

Rule two: the stop rule is written once, not per user

Dots' safety model is good and it is in the wrong place. Allow, block, or require approval per action is a stop rule, and it ships as a personal setting. Every employee will write their own, most will not, and the defaults will decide.

I published The Stop Rule yesterday about two vendors that shipped self-repairing agents in the same week. Salesforce's Agent Designer limits self-healing to two attempts and a $3 ceiling, then stops and escalates. The other published a loop that runs until a fix passes. My own autonomous pricing agent had no stop rule and sat 14% under benchmark for two quarters.

For a company running Dots, the stop rule is a product decision made once, org-wide, and pushed down: anything that touches money, contracts, individual customers, or decisions that compound over time requires a human approval that a user cannot switch off in their own settings. The admin toggle for the enterprise beta is where that rule goes before the first dot is enabled, not after the first bad week.

Rule three: the rules are the interface, so put a designer on them

The coverage is about the bubble. The bubble is not the design surface. The approval prompt is. The line between read-only research and acting is. What the dot shows you when it was wrong, how it discloses what it read, and how it repairs trust after a mistake are.

That is the whole argument of Design Just Got Promoted: drawing got cheap, deciding did not, and design's job moved from producing artifacts to specifying behavior, with the wrong path owned as a first-class surface. Dots is the cleanest case yet. A product team that treats the dot as an OpenAI feature and not as a surface they are responsible for will find out what their employees approved by reading the incident report.

The FDE angle: a specialist dot is the deliverable, not the promise

Specialist dots with company credentials, piloted on procurement and invoicing and contracting, sold through Agent 365, are the forward deployed engineer's deliverable offered direct. If a dot lands the procurement workflow in a week, the rule in Building the FDE Team: Reporting Line, Bar, Pairing, and Comp to size the team to shrink per deployment just got a deadline.

What the dot does not remove is the promise. In In the Room, Not on the Quota: Who Owns the FDE Promise the instrument is a one-page handshake before the signature: every outcome in the proposal sorted into ships, builds, or no, with sales owning the number and the builder owning the columns. A specialist dot pilot needs the same page. "Handle everything" is a promise nobody owns, and "always review consequential work" is a job nobody has been assigned. Assign it, on paper, before the pilot starts, and pair the person who sets the dot up with the person who owns adoption, because they are different skills.

The number that did not launch

Agent count is a vanity metric. A dot per user is agent count with a cloud computer attached. The number that matters, the one the Product Leadership argument keeps coming back to, is which coordination roles stop being necessary and which new ones appear. That is answered by one measurement: how many dots still complete production work after ninety days, and what each completed outcome costs.

Nobody launched with that. Ask for it before you approve the second dot.

What to do this week

If your company is on Business Premium or has an enterprise admin who can flip the beta, do three things before Friday. Write the owner rule and the stop rule on one page. Put a designer on the approval and recovery surfaces. And if a specialist dot pilot is on the table, draft the ships, builds, or no page with the vendor before anyone signs.

Related answer: Who should own an OpenAI dot inside a company?

Sources: OpenAI launches dots, always-on AI agents with their own cloud computers, Ana Maria Constantin, The Next Web, September 29, 2026. OpenAI launches Dots, its bubbly agentic avatar, Lucas Ropek, TechCrunch, September 29, 2026. 39 PM AI Agents Deployed: What Stuck, What Died, and Why, falkster.com.