Your AI Agent Sleeps 8 Hours a Day. Dots Never Clocks Out.
In this article
Bottom line: On September 29, 2026, OpenAI launched Dots at DevDay. Dots are "always-on" agents powered by GPT-6 Astra.
Each one gets its own cloud computer, connects to 4,000+ apps and a browser, and keeps working on your goals after you close the tab.
They're available to ChatGPT Pro and Business Premium users, with enterprise "expert" Dots in testing.
I haven't used Dots yet, so this piece covers what's confirmed, what Hacker News is arguing about, and the test I'd run before trusting one with real work.
I Haven't Tested It. Here's What I'd Do Instead.
I'm going to break format here. Normally I'd hand you a spreadsheet of 47 tests and a dramatic winner. Dots has been public for about a day, and I don't have access.
Anyone posting stopwatch numbers right now is guessing or making them up.
What I can do is read the launch coverage, read the arguments, and write the test plan I'd run once I get access. That plan is useful either way, because "always-on" changes what you're evaluating.
What Dots Actually Is
Per OpenAI's announcement and 9to5Google's write-up, a Dot is an agent that runs on its own cloud computer.
It can use a browser and connected apps to research, draft documents, write software and handle other tasks for you. It also learns from your feedback over time.
Coverage says it connects to more than 4,000 apps and takes text or voice instructions.
Two details matter more than the feature list. The first is that it isn't tied to your session. A chatbot waits for you to type.
A Dot keeps working toward a goal while you're asleep, in meetings, or on a plane.
The second is where it lives. According to CBS News and Slashdot, Dots can plug into Slack and Teams channels. There they can help brainstorm, update calendars and review customer feedback.
That moves the agent out of a private chat window and into the place where your team talks.
Availability is narrow for now. It's ChatGPT Pro and Business Premium, with enterprise users testing dedicated "expert" Dots. Several outlets frame it as a response to Meta's recent Muse release.
Why "Always-On" Is a Different Product
When I evaluate a chatbot, I care about answer quality. I ask it something, judge the reply, and move on. The failure mode is a bad paragraph, and I can see it immediately.
An always-on agent fails differently. The cost of a mistake now depends on how long it ran before you noticed. A wrong answer costs you a minute.
A wrong action repeated for six hours costs you a mess to clean up.
That's why I think the usual benchmark talk misses the point. The right questions are about supervision:
- How does the agent tell you what it did?
- How easy is it to stop?
- What can it touch without asking?
Gizmodo's headline captures OpenAI's framing: with Dots, OpenAI wants you to stop being afraid of its AI agents. Reassurance is a product decision here, not a side note.
A company doesn't ship "stop being afraid" messaging for something with no fear attached.
What Hacker News Is Fighting About
The Dots thread passed 500 points and several hundred comments within hours, per the daily digests. The split, as summarized in those digests, runs along a few lines.
Privacy architecture. If the agent has its own computer and holds your credentials, where does that data live, who can read it, and what happens when you delete it?
"Always listening" came up repeatedly, which is a fair worry once voice is in the mix.
Innovation or surveillance creep. One camp sees the next interface after chat. The other sees a persistent, permissioned observer sitting inside your work tools. Both can be right.
The joke layer. TechCrunch reports that the internet is convinced xAI trolled the Dots launch. That's gossip and I won't dwell on it, but it shows how crowded and combative this category has become.
I'd add a fourth issue that I saw less of. The hard problem isn't whether the agent is smart. It's whether you can stay a competent supervisor of something that never stops.
The Test I'd Run: Five Days, Five Questions
Here's my plan for when I get access. Steal it.
Day 1: The Boring Task
Pick a task you already do weekly and know cold, like a competitor roundup or an inbox triage. Do it yourself and time it. Then hand it to a Dot with the same instructions.
Score two things: accuracy against your own version and your review time. If checking the Dot's work takes 80% as long as doing the task, you've bought nothing.
Day 2: The Unattended Overnight
Give it a multi-step goal at 6 p.m. and don't touch it until morning. Then read the activity log cold, as if a coworker left it.
Ask whether you can reconstruct what happened in under five minutes. If you can't, the agent may be capable but it isn't supervisable, and that matters more.
Day 3: The Ambiguity Trap
Give it an instruction with a deliberate gap, like "clean up the shared folder." Watch whether it asks a question or guesses.
I care less about the mistake than about the guess. An agent that stops and asks is one I can trust with more. An agent that fills gaps confidently is one I have to fence in.
Day 4: The Kill Switch
Interrupt it mid-task. Revoke one app connection. Change the goal halfway through. Time how long each takes to actually take effect.
"Always-on" is only comfortable if "off" is instant and unambiguous.
Day 5: The Blast Radius Audit
List every app it touched that week. For each one, ask if you'd have given a new intern that access on day one.
Most people will find they granted broad access in a hurry because the setup screen made it easy. Convenience quietly sets the permission policy.
How I'd Score It
I'd track five numbers and put them in a table you could hand to your team:
| Dimension | What I'd measure |
|---|---|
| Accuracy | % of outputs matching my own version |
| Net time saved | Task time minus review time |
| Auditability | Minutes to reconstruct a night's activity |
| Ambiguity handling | Asked vs. guessed, out of 10 traps |
| Stop latency | Seconds from "stop" to actually stopped |
Whatever the results are, I'll report them with the misses included. I've read enough launch-day threads to expect that the demo tasks will look great and the messy real-world ones will be the story.
Who Should Care Right Now
If you're on Pro or Business Premium, run the five-day test on low-stakes work before anything touching customers, money, or shared drives.
Start with read-only permissions and widen them only when the logs earn it.
If you manage a team, the Slack and Teams integration is the part to think hardest about. An agent in a shared channel has a different social footprint than one in your private tab.
Decide the norms (who can assign it work, whose instructions win) before someone discovers them by accident.
If you're waiting on enterprise, watch the "expert" Dots pilot.
Purpose-built agents with narrower scope may prove safer than the general one, though that's my inference, not something OpenAI has demonstrated.
If you're skeptical of the whole category, that's reasonable. Nobody has long-run evidence on an agent that learns from your feedback for months. That includes me, and it includes OpenAI.
The Thing I Can't Stop Thinking About
The framing "your AI agent sleeps and Dots doesn't" is catchy, but it hides the real shift. Agents that run around the clock change you from a user into a manager.
Managing is a skill, and most of us have never had to practice it on something that works while we sleep.
The best outcome isn't an agent that never makes mistakes. It's one whose mistakes are small, visible, and reversible.
That's a thing you can test for, so I'd rather see a Dot fail loudly in week one than succeed silently until week six.
Have you handed any real work to a background agent yet, and what was the first thing it did that you didn't expect?
Sources: OpenAI, 9to5Google, CBS News, Gizmodo, Slashdot, TechCrunch
