ChatGPT Dots: how to use an always-on AI agent without drowning in average work

OpenAI launched ChatGPT Dots on 29 September 2026. Dots are agents that keep working when you close the laptop. We are in the UK, where Dots aren't available yet, so we recorded a first reaction instead of a review.

After the video, we checked our own claims against the launch coverage and the research on delegation. 2 things changed. The OpenClaw connection is clearer than we said on camera. And our worry about "delegation junk" already has numbers behind it.

What is ChatGPT Dots?

Dots are always-on agents inside ChatGPT. According to SiliconANGLE's launch report (29 September 2026), GPT-6 Astra powers them, and each dot works from its own cloud computer that you can open to watch what it is doing. A dot can connect to more than 4,000 apps through OpenAI's plugins, run several projects at once, and learn your preferences from feedback. You can message it or call it in ChatGPT, or reach it in Slack and Teams.

The first dot comes with ChatGPT Pro and Business Premium at no extra cost. Chatting with a dot doesn't count towards your limits, but The Register (29 September 2026) reports that tasks it starts in Codex or ChatGPT Work count as normal. Extra dots will cost more, at prices OpenAI hasn't published.

ChatGPT Dots vs Codex vs ChatGPT

ChatGPT waits for your next message. Codex takes a task and comes back with a result. Dots sit between the 2: you give a goal, and the agent decides the next steps, keeps going overnight, and reports back.

In practice, you stop managing each step. You manage the goal and the review.

ChatGPT Dots vs OpenClaw and Grok Bot

On camera we weren't sure how OpenClaw's creator fits in. Here is the record. Peter Steinberger, who built the open-source agent OpenClaw, joined OpenAI in February 2026 (TechCrunch, 15 February 2026). Sam Altman said his job was to "drive the next generation of personal agents". OpenClaw moved into an independent foundation that OpenAI sponsors. So Dots come from inside OpenAI, with the person who made this category popular on the team.

We use xAI's Grok Bot every day (our setup is here), and it already has its own computer and runs overnight. On paper, the differences are the 4,000+ app connections, the Slack and Teams access, voice calls, and preference learning. Whether those change daily work is something only a test will show. One thing neither product gives you is a choice of model per agent. In our internal agent app, we run 1 agent on Fable 5 and another on Haiku, depending on the job.

Why is ChatGPT Dots not available in the UK?

OpenAI said Dots "will not reach Pro users in the European Economic Area, Switzerland or the UK for now" (Trending Topics, September 2026). It gave no reason. The likely friction is data protection: an agent that reads your email and calendar 24 hours a day needs a clear legal basis under GDPR. OpenAI's Operator agent and Sora followed the same pattern and launched outside Europe first.

The delegation risk has a name: workslop

Our main worry in the video was delegation junk. Researchers call it workslop. A 2025 survey by Stanford's Social Media Lab and BetterUp Labs, published in Harvard Business Review in September 2025, asked 1,150 US full-time employees about AI-generated work. 40% had received workslop in the previous month. The people receiving it spent nearly 2 hours fixing each case.

Speed doesn't rescue it either. In a randomised trial by METR (July 2025), 16 experienced open-source developers took 19% longer to finish tasks with AI tools. Afterwards, they estimated the tools had made them 20% faster. Feeling productive and being productive are 2 different measurements.

An always-on agent multiplies both effects. It produces while you sleep, so it can produce average work at a volume no person can review.

Why "allow access" stops protecting you

Every Dots demo is a chain of "allow access" buttons. Repetition changes how we read them. In 2013, Devdatta Akhawe and Adrienne Porter Felt studied more than 25 million browser warnings and found Chrome users clicked through the SSL warning 70.2% of the time (USENIX Security 2013). In 2016, Brigham Young University researchers used brain scans to show that the brain's response to a security warning drops each time it sees the same one again (Journal of Management Information Systems).

OpenAI builds in some checks. SiliconANGLE reports that actions touching your accounts go through auto-review, and research mode is read-only. OpenAI's own advice is to review consequential work because agents "can still make mistakes". The last check is still you, and you are the check most likely to be on autopilot.

How to delegate to an always-on AI agent

1. Write your priorities down first. One document that says what you care about, what you ignore, and how you like things written. Point every agent at it. Our agents read our company second brain before any task.
2. Start with read-only work. Research, summaries, news filtering. No sending, no posting, no paying.
3. Grant access 1 app at a time. Write down what you allowed and why. If you can't say why, don't allow it.
4. Delegate the draft. Keep the decision. The agent gathers and prepares. You choose.
5. Ask for a fixed morning report: what it did, what changed, what it needs from you. 3 short sections you can read in 2 minutes.
6. Read every piece of work before anyone else sees it. If you wouldn't sign it, don't send it.

Our first read

We haven't used Dots, and we will change our minds once we do. What we already know is that the agent never gets tired, and we do. The value of an always-on agent depends on the person who writes its brief and reads its report, so that person has to stay awake.