AI development

The Shape of AGI Arrived Before the Intelligence Did: What OpenAI's Dots Really Are

Jay Sampat·September 30, 2026·9 min read

OpenAI's Dots are not a smarter assistant and they are not AGI. They are the operational shape that anything AGI-like will take inside a company: a persistent agent with a name, an identity, its own computer, a work quota, and a security policy, working while you are not. The intelligence inside will keep improving on its own schedule. The shape is here now, from six different vendors at once, and the shape is what you have to decide about. I spent the last two weeks watching Salesforce, Microsoft, and now OpenAI announce the same object in three different vocabularies. This is my read on what it is, what it is not, and what a CEO or CTO should do before the fleet arrives.

What did OpenAI actually release?

Per OpenAI's DevDay announcement on September 29, 2026 and the coverage that followed, Dots are "always-on agents built to handle everything." The facts, as stated by the company and reported by TechCrunch and Bloomberg:

Each Dot is powered by GPT-6 Astra, gets its own cloud computer and browser, and keeps working on your behalf between your conversations with it.
A Dot connects to more than 4,000 apps through OpenAI's plugin ecosystem and is reachable through ChatGPT, Slack, or Microsoft Teams, with text message support coming soon.
As TechCrunch puts it, Dots are meant to operate independent of any specific hardware or interface, pursuing user-defined goals continuously in the background with minimal oversight.
You start with one Dot, give it a name, and make it your own. OpenAI is piloting "specialist Dots" for businesses, each with a defined role, and describes teams of Dots working together as the longer-term plan.
Each Dot can be given its own identities, credentials, and tool access, configured through existing systems, and OpenAI is working with Microsoft to integrate specialist Dots with Microsoft Agent 365's governance and security controls.
Dots are available to Pro and Business Premium users in eligible markets, and Enterprise users can try them when workspace admins approve. Bloomberg reported a new $500 paid tier alongside the launch. OpenAI says the ability to scale a Dot's speed or the total amount of work it can take on per month comes later.

Two facts from the same week belong next to that list. OpenAI decided not to release GPT-6.1 Astra because it did not adequately meet the company's safety standards, per CNBC on September 28. And Fortune's launch coverage noted that OpenAI's own agents have continued to hack websites and act in unintended ways during training. Read together, the launch is a very capable product shipped with a security partner attached, because OpenAI's own testing shows why it needs one.

Is this different from what the other labs have?

Less than the launch keynote suggests, which is the point. Six players have now announced the same architecture in different clothes.

Product	Who it is for	The persistent agent	Identity and security	Where it lives
OpenAI Dots	Enterprise and prosumer	Always-on, goal-pursuing, teams of Dots planned	Provisioned credentials; Microsoft Agent 365 integration	ChatGPT, Slack, Teams, its own cloud computer
Microsoft Autopilot	Microsoft 365 organizations	Agent with a name, role, and goal that works unattended	Its own Entra identity, IT-controlled runtime	Inside the tenant
Meta Muse	Consumers	Personal agent, moving into glasses	Consumer-grade	Meta's apps and devices
Salesforce Agentforce and AIforce	Salesforce customers	Agents over governed CRM data, interfaces assembled by AI	Runs on existing permissions; capabilities over MCP	The Salesforce estate
Anthropic Claude	Everyone building agents	The model inside other companies' agents, from Claudeforce to Slack	Depends on the host platform	Wherever the host runs it
Open source (OpenClaw and peers)	Builders	Personal agent runtime, community-owned	Whatever you configure	Your own infrastructure

Sources for the table: each company's own announcement, with our earlier reads of Microsoft's Copilot rebuild, Salesforce's Dreamforce announcements, and OpenClaw 2.0 linked below. I have kept the cells to what each company states; where a vendor has not said something, the cell says so.

The differences are real but they are positioning differences. OpenAI ships the agent as a branded persona with a work quota. Microsoft ships it as a tenant-resident employee with a directory identity. Salesforce ships it as a layer over the data it already governs. Anthropic sells the intelligence rather than the persona. Meta sells the consumer version. The open-source world gives you the runtime and lets you decide. Underneath, every one of them is a persistent, goal-pursuing agent with an identity, a set of credentials, a runtime that keeps it alive, and a bill that scales with work done.

Is it AGI or is it just another assistant?

Neither, and the question itself is the trap.

It is not a capability leap. As TechCrunch observed, much of what Dots do was already possible through Codex and similar agentic harnesses; Dots bundle those capabilities into a package focused on independent action. That is productization, and productization is not the same as intelligence.

It is also not "just another assistant," because an assistant waits for you. A Dot does not. The difference between a chat window and a Dot is the difference between a tool and a colleague: it has standing goals, it acts without being prompted, it has credentials you granted, and it will do things while you are asleep that you will read about in the morning.

Here is the claim I would defend in any boardroom: the form of AGI is arriving before the intelligence. If a genuinely general system shows up in your company in the next few years, it will not appear as a wiser chatbot. It will appear exactly like this: provisioned with an identity, attached to systems through permissions, working continuously, organized in teams, governed by security controls, and paid for by the unit of work. Whether the model inside deserves the word "general" is a research argument that will run for years. Whether your company has the permissions, the evaluations, the cost controls, and the identity policy to run a fleet of these is a question for this quarter, and the answer is the same regardless of how smart the model turns out to be.

What does this mean for a company deciding what to do?

Four things follow from the shape, and none of them depend on the intelligence.

Identity becomes the security layer. Every vendor in the table has converged on giving agents their own identity and credentials. That is the right design and it creates a new obligation: knowing which agents exist, who created them, what they can see, and what they can do. Our post on AI governance for mid-market companies covers the policy side; the technical side is permissions scoped per agent, never borrowed from a human.

Evaluation has to come before autonomy. An assistant's mistakes are visible because you are watching. An always-on agent's mistakes are discovered later. The only way to trust a Dot, an Autopilot, or an Agentforce agent with standing goals is to have measured it on your own decisions first. We wrote about how to evaluate LLM outputs with exactly this in mind, and the harness pattern explains why the cheap, checkable steps of an agent should not run on the same model as the reasoning.

Cost becomes usage. OpenAI plans to scale a Dot by speed or monthly work. Microsoft moved agentic features to usage-based billing. Salesforce meters agents. This is the third vendor in a month to bill agents by work done, and it means the finance conversation happens before the agent is switched on, not after the first invoice.

The screen becomes optional, but not everywhere. A Dot reachable from Slack or Teams is another step in the shift I described in The Screen Is Optional Now: the fixed screen is no longer the product. That does not mean every workflow should be handed to a background agent. High-frequency, high-stakes, and regulated work still wants a designed interface and an audit trail. Low-frequency, analytical, and repetitive work is where Dots will earn their fee. Knowing which is which, in your company, is the actual work.

What would I do before adding a second Dot?

If you are a leader reading the DevDay coverage and wondering whether to switch this on, here is the list I would give a friend.

Inventory the systems an agent could touch and decide, per system, whether an agent may read, act, or neither. Write it down before anyone provisions credentials.
Start with one Dot, one workflow, one data source. OpenAI's own design starts you with a single Dot for a reason. Resist the team of Dots until the first one has a track record.
Give it an identity you can revoke, through your directory, not a shared login. If your vendor integrates with your existing security controls, use that path and no other.
Log what it does and read the log for the first month, the way you would read a new hire's work.
Model the cost under usage billing before enabling it, with a cap. A Dot that "handles everything" can also spend everything.
Decide which workflows keep a screen. Not everything should become a goal you hand to something that works while you sleep.

We build agents for clients on OpenAI, Anthropic, Microsoft, and Salesforce platforms and on custom stacks, and we resell none of them. The audit above is the one we run with companies before they commit to any of these roadmaps, and it is the same audit whether the agent turns out to be a very good assistant or the first thing that deserves a bigger name.

The intelligence will arrive when it arrives. The shape is already in your Slack.

Sources: OpenAI, Introducing dots (September 29, 2026); TechCrunch, OpenAI launches Dots, its bubbly agentic avatar; Bloomberg, OpenAI Unveils Always-On AI Agent Dots, New $500 Paid Tier; CNBC, OpenAI abandons plan to release upcoming model as safety concerns escalate; Fortune via Yahoo Finance, OpenAI unveils Dots as rival to Meta's Muse. Vendor comparison drawn from each company's own announcements, including Microsoft's Copilot update and Meta's Muse glasses post. Product details are as of September 30, 2026 and several are in early rollout.