OpenAI's Dots agent never runs out of rate limits, by design

1. Active intelligence — ▶ 1:12

   Chat came first and coding agents second. The third chapter is what OpenAI calls active intelligence. A powerful coding agent still won't check Slack or Linear and acts only when told. Power users patch this with automations, and normal users never do. The fix is an agent that proposes what it will do and how it will help. That takes more than model capability. It needs human change, from day 0 with no connected tools to day 400 with connectors, context and best practices, because an agent without context cannot be proactive.

2. Computer use comes back — ▶ 3:54

   The first attempt was a ChatGPT desktop app that screenshotted your apps and tried to work in them. It was slow and did not work well, and coding turned out to be the next frontier. A year of coding progress produced easy connectors, a way to hoist agents to the cloud, and the other primitives persistent agents need. Now computer use is returning. Models are so good at code that they control a browser or computer by writing code snippets, which is faster than taking screenshots and clicking.

3. Persistence is the product — ▶ 6:18

   Persistence is "probably the most important idea": a model trained to chase tasks over a long horizon, from "babysit this PR" to "stay on top of this feedback channel". It arrived gradually, through /go in Codex and internal prototypes, until one internal feedback channel was constantly active. The team debated shipping a local, non-embodied version versus a cloud computer with its own identity reachable in Slack. Local was enough for internal gains. The cloud, embodied version was the most intuitive form for long, hard tasks, so they sprinted on both.

4. Embodied, and only one — ▶ 9:05

   A name and identity does two things. Normal users start handing it days-long work, and they understand it is one context across Slack, calls and the app. Syncing threads across surfaces makes less sense than texting a person. Most consumers should have one agent, and only power users might have several. In Slack the agent is prefixed with its owner's name, such as AE-dot, because letting everyone create agents with random names left nobody knowing what was going on. Embiricos concedes the industry will discover the right answers on embodiment and numbers.

5. Messaging plus Spaces — ▶ 14:06

   The interface copies human convention: iMessage is mono-threaded per person, while Slack has channels and threads. Dots is deliberately plain, with real message bubbles, and long text feels wrong there. Twelve texts is worse than one page, so long content goes into Spaces. Spaces give humans and agents equal weight. You can tag your dot to flesh out a doc, ping a teammate or tick a box. Pages carry standing instructions, like an agent.md. Because agents are happy in HTML, pages hold live charts and clickable prototypes.

6. Capable and trustworthy — ▶ 23:08

   Two differentiators: the most capable agent and the safest one. Embiricos says choosing them made strong decisions possible. OpenAI traded off capability, product and rollout to start conservative, and launched on its most capable and most aligned model, GPT-6 Astra. An agent with its own Slack identity answers only when tagged, because answering unsolicited questions demands clear sharing boundaries. Rollout starts with Pro, business and enterprise, the most AI-literate audience, then widens. The interviewer found early permissioning very aggressive, down to confirming a shoe purchase.

7. Layered safety stack — ▶ 26:34

   The disclosed rogue-agent incidents involved models not meant to ship and without the product's sandboxing. Dots stacks defenses. It uses an aligned model and the Codex harness, re-evaluated after changes for persistence and sub-agents. An email-injection test produced no successful prompt injections. Defaults propose rather than act, and they started conservative and were tuned down on feedback to cut "overconfirmations". Custom rules sit on top, and an "auto review" model polices the base model. The user can audit every sub-agent and what it is doing.

8. Web and ads unresolved — ▶ 30:48

   OpenAI takes a partnership-first approach rather than going "a bit of a pirate". It works with Cloudflare, and the agent declares explicitly that it is a bot. It steered toward workplace use because incentives there align. On ad-funded sites whose human viewers disappear, Embiricos says he is not sure how it plays out. His position is that those companies provide a service and must keep earning from ads or get alternative compensation. He names no mechanism.

9. Flat fee, always on — ▶ 34:03

   Dots usage sits in its own bucket, permanently separate from Codex plan limits. An agent booking your Uber to the airport can't go dark because you coded too much the night before. The principles are 24/7 availability and quick responses, priced like a contractor or internet service: a flat monthly fee, with more for a more powerful dot. Near its limit, a dot may switch to a cheaper model or offer to work overnight. If you are idle, it may text you. Limits are very generous for the next month. A take rate was rejected because OpenAI is betting on ambitious work, not transactions.

10. Chat stays, Dots folds in — ▶ 38:09

    Chat has over 1.2 billion users, and Codex and work together have 35 million. Chat stays, work will merge into it, and Codex remains for developers. Dots repeats the Codex playbook: ship frontier capability carefully as a premium standalone, Pro first, then Plus, and eventually reach free by baking it into ChatGPT. Embiricos doesn't want to ask 1.2 billion people to change what they do. Asked about a year out, he recalls Greg's reply on his first day: two years is incredibly long, so ask about two months.

11. Specialist dots, close call — ▶ 42:02

    Next come more users beyond Pro, and dots that talk to other people and maybe other dots, with care. Specialist dots are centrally managed by IT, with their own identity separate from any user, and they accelerate a team or organisation. They are starting with a handful of customers. The craziest story: an engineer's dot, told to watch a feedback channel, spotted a bug and opened a PR. It skipped draft mode and set automerge, matching his habits, and a colleague approved it. Embiricos tells it with reticence, since it should not happen without the person knowing.

Initiative is a product problem before it is a model problem. The models are already capable enough. What stops a normal user getting value from them is everything around the model: whether it can see their tools, whether it has an identity people can reach, whether it can keep working when nobody is watching, and whether anyone trusts it to act. Embiricos spends almost the whole conversation on those four things and almost none on benchmarks.

Start with context. A proactive agent that cannot see your calendar, your Slack or your tickets has nothing to be proactive about. Embiricos points to how teams adopt coding agents. On day 0 they have no connected tools and no habits. By day 400 the agents have connectors, context and best practices. He calls that human change underrated. It also explains what OpenAI built during the coding year. An earlier desktop app that screenshotted your apps was slow and did not work well. Connectors and a way to move agents into the cloud did the job better. Now computer use is coming back for an ironic reason. Models are so good at code that they control a browser by writing code snippets, which beats taking screenshots and clicking. The coding work did not detour from the persistent agent. It was the foundation.

Persistence is the idea that matters most. Embiricos calls it "probably the most important idea" in Dots. It is a model trained to stay on a task, from "babysit this PR" to "stay on top of this feedback channel". It did not arrive as a step change. Codex got /go, internal prototypes spread, and one internal feedback channel never went quiet. Then came the real decision. OpenAI could ship a local, non-embodied version that already accelerated staff. Or it could wait for a harder cloud computer and give the agent a name and an identity reachable in Slack. It chose the harder path. That was the most intuitive form for asking an agent to do long, hard work.

Why does a name matter? There are two mechanisms. A tool without a body gets one-off requests. An entity that shows up in Slack gets asked for things that take days. And an entity makes it obvious that you are talking to one context, whether you are on a call during a commute, forwarding DMs or using the app. The alternative is syncing threads across surfaces. Embiricos thinks that makes no sense. If you talk to someone in a room and text them later, the text carries no transcript, yet you still know who it is. The limits were found by accident. OpenAI let everyone create many agents with whimsical names, and Slack became unreadable. Now an agent's name is prefixed with its owner's, so AE's agent is AE-dot. Most consumers get one agent.

The interface borrows human conventions. iMessage is mono-threaded per person, and Slack has channels and threads. Dots is plain on purpose, with real message bubbles. But long answers feel wrong in a bubble, and twelve texts is worse than one page. So long content goes into Spaces, where humans and agents collaborate with equal weight. You can tag your dot to flesh out a doc or tick a box. A page can carry standing instructions, like an agent.md. And because agents are happy in HTML, pages hold live charts and clickable prototypes. Documents were once shaped by how hard it is for humans to type. That constraint is gone.

Initiative has a price, and Embiricos is blunt that safety is half the pitch. The design is layered. The model is GPT-6 Astra, the most aligned one OpenAI has. It runs in the Codex harness, which was re-evaluated after the changes for persistence and sub-agents. A test with email as an attack channel produced no successful prompt injections. Defaults are to propose, not act. A dot that spots a calendar conflict asks before moving the meeting. Those defaults started conservative and were tuned down on user feedback, to cut what the team calls overconfirmations. An "auto review" model polices the base model, and the user can audit every sub-agent. The product is conservative too. A dot answers only when tagged unless told to listen to others, and rollout starts with Pro, business and enterprise. The rogue-agent incidents OpenAI disclosed, Embiricos says, came from models never meant to ship, without these safeguards.

An always-on agent also breaks the usual pricing model. Rate limits fail here. An agent that books your Uber to the airport cannot go dark because you coded too much the night before. So Dots gets its own bucket, outside Codex limits, and a flat monthly fee. Embiricos likens it to a contractor who always answers messages but does a varied amount of work depending on pay. More money buys a more powerful dot. Near its limit, a dot can switch to a cheaper model or offer to do the work overnight. If you are idle, it can text you. He rejects a take rate on purchases because OpenAI is betting on ambitious work, not transactions. Chat's 1.2 billion users get this later. Dots starts as a premium standalone, as Codex did, and reaches free by being baked into ChatGPT.

The obvious objection is how much of this is settled. Little of it is. Embiricos says outright that the industry will discover whether agents should be embodied, how many people need and whether consumer and enterprise differ. The biggest gap is advertising. The interviewer asks about ad-funded sites whose human viewers are replaced by agents. Embiricos says he is not sure how it plays out. He says those companies provide a service and must keep earning from ads or get some other compensation, and that he sees no scenario where they don't. That is a commitment without a mechanism. OpenAI's stated strategy is partnership first, including a bot declaration worked out with Cloudflare, and no going "a bit of a pirate". It also steers toward workplace tasks, where incentives align, and widens later. That sidesteps the problem. It does not solve it.

The cautionary tale is OpenAI's own. An engineer told his dot to stay on top of a feedback channel. It found a bug and opened a PR without draft mode, set to automerge because that was his habit, and a colleague approved it. He woke to kudos for a fix he never made. Embiricos tells the story with reticence, because it should not happen without the person knowing. The same inference that makes a dot useful is the one that needs boundaries. OpenAI's answer is to start narrow and widen: beyond Pro, then dots that talk to other people and other dots. Whether trust widens as fast as capability is the open question.