Sign in Subscribe
news

2026-10-03 OpenAI Dots - always-on agents, with a memory you can't inspect

Sebastien Dubois

Sebastien Dubois

— 6 min read

Canonical version: 2026-10-03 OpenAI Dots - always-on agents, with a memory you can't inspect.

On Tuesday (29 September), at DevDay, OpenAI launched Dots: always-on agents that live inside ChatGPT and keep working when you close the window.

Each dot runs on GPT-6 Astra, has its own cloud computer and browser, connects to the apps you've plugged into ChatGPT (more than 4,000, says OpenAI), and learns your preferences over time. You reach it in ChatGPT, Slack or Microsoft Teams, by message or voice call. And it reaches out to you too, with progress, questions and decisions it needs from you.

The first reaction on Hacker News compared it to OpenClaw as a managed service. OpenClaw's creator, Peter Steinberger, joined OpenAI in February.

What shipped

  • A primary dot per user, with a name, an avatar (or a pet) and a handle like @yourname-dot
  • Background work. While you're away, the dot reads your connected apps looking for ways to help ("proactive research"), runs scheduled checks and reminders, and juggles several projects at once
  • Its own computer, optionally yours. You can open the dot's cloud computer to inspect its work. Access to your laptop is off by default
  • Delegation. It can start tasks in OpenAI Codex and ChatGPT Work
  • Specialist dots (enterprise pilots): agents with their own identity, credentials and hardware for one job, like procurement or customer support. OpenAI is working with Microsoft to manage them through Agent 365

OpenAI's examples: a dot investigates a bug the moment it's reported in Slack, turns a design into a working app, opens pull requests with videos of the change. The announcement also shows five "extension of you" scenarios, each with a named dot: a developer turning feedback into tested fixes, a launch lead revising a launch, a scientist updating an analysis, a sales lead updating a proposal, and a content creator automating content production with a dot that learns their voice. That last one is clearly aimed at people like me. In a smaller example, an early tester's dot noticed he had forgotten to invoice a publication, prepared the invoice, and sent it once he approved.

Who gets it, and what it costs

  • Pro users, from 29 September, in markets excluding the European Economic Area, Switzerland and the UK. Pro costs $100 to $500 per month
  • Business Premium in all supported regions; Enterprise/Edu/Healthcare as an admin-enabled beta
  • Nothing for Free, Go or Plus users. Nobody under 18
  • The first dot is included in the plan. Chatting with it doesn't count toward your usage limits; tasks it starts in Codex or ChatGPT Work do

Tibo Sottiaux announced Dots on X as "Included in your Pro plan, without drawing down on any of your usage". The post got an X Community Note, and he replied the next day: "I got community noted, but the note is wrong." His version: work the dot does directly uses none of your allowance, a Codex task it creates draws usage as usual, more speed and bandwidth will become a paid extra, and "the baseline functionality will always just be included in your plan." OpenAI's published terms are narrower: extended limits are promised for the first month only. On the same day, OpenAI also announced that the $200 Pro plan's Codex and Work allowance will drop from 20x Plus to 10x on 30 October. A dot that decides to hand work to Codex spends from a smaller budget, and OpenAI hasn't said whether it will warn you first.

Safety design

Background research uses read-only tools, enforced in code: it can't send messages, change app content or control a browser. Before a dot sends an email or changes a file, a separate model (Auto-review, borrowed from Codex) checks the action against your instructions, your Custom Rules and the safety requirements, from outside the environment the dot can touch. Passwords go through a credential service and never enter the model's context. Changing a password or moving money always comes back to you.

The GPT-6 Astra system card reports the limits. When OpenAI chained five intervening tasks with shifting (and unstated) permissions, 8.6% of samples got flagged for scope problems. With ten tasks, 19.7%. None were high-severity, but the rate rose with the number of tasks. If you build long-running agents yourself, re-state the scope between tasks instead of setting it once.

The memory problem

A dot builds a picture of you from your conversations and connected apps. According to OpenAI's own FAQ, you can't view, edit or delete individual dot memories. You can only reset the whole dot, and disconnecting an app doesn't remove what it already learned. That bugs me.

I keep my AI agents' memory in plain Markdown inside my Obsidian vault, exactly so I can read it, fix it and take it elsewhere (AI Agent Memory, AI Agent Portability). I won't use an assistant that knows me but can't show me what it knows. Several HN commenters made the same point from the lock-in angle: once an agent holds your integrations, history and memory, switching gets expensive.

Hacker News reactions

The launch thread hit 765 points and 646 comments, and the mood was mostly negative:

  • "What is this?" Many couldn't figure out the product from the page. "Hosted OpenClaw" and "Grok Bot 2.0" were the common shortcuts, and Clippy came up a lot
  • Name and mascot. Lots of jokes about "dots" and the cute characters, and suspicion about why always-on agents get cuddly designs
  • Trust. People asked who would give an agent write access to their life. One commenter quipped that their personal assistant can now stop working if the company disagrees with what they're doing
  • Price and region. $100/month minimum (vs Meta's Muse, which has a free tier), and no Pro access in the EEA or the UK
  • Developers don't need it. A recurring comment: "my throughput is limited by my approval". An agent working overnight doesn't help if you're the bottleneck
  • Early hands-on reports were rough. One user graded it "B-" after the dot asked for confirmation again and again before booking a shuttle. Another said the cloud machine died several times while working on a GitHub issue, and the dot then went silent overnight
  • One technical finding: a commenter reported that the dot's cloud environment intercepts HTTPS with an OpenAI-issued certificate (unverified)

Not everything was negative. A heavy Grok Bot user explained why separate, domain-specific always-on agents are useful. And one commenter who had already set up OpenClaw said he was excited to get a managed always-on agent instead.

On X, Sottiaux's launch post passed 2.9M views and earned that Community Note. Ben Thompson's Stratechery piece is paywalled; the teaser says: "OpenAI's Dev Day showcased a product that is, frankly, pretty confusing. However, there is more vision here than it might seem."

Outside HN, Ethan Mollick sees Dots and Meta's Muse as part of a wave of "Clawlikes", and says his own personal agents now catch his mistakes. Amplifying.ai had Dots build the same app 8 times with the stack left open: 8 out of 8 landed on Cloudflare Workers, D1 and ChatGPT sign-in through ChatGPT Sites.

My take

All the major labs now have always-on agents: Meta has Muse, xAI has Grok Bot, Google has Gemini Spark, and Anthropic folded Claude Cowork into the Claude app, with Claude Tag in Slack. Dots is OpenAI's entry, and its safety documentation is unusually detailed.

But I care about owning my context. Until I can see and edit what my agent believes about me, I'll keep building my agents on top of my own notes. And here in Belgium, the question is academic for now: Pro access excludes the EEA.

If you can try it, start with read-only work (research, monitoring, briefings) and grant write access one action type at a time.

The full breakdown (features, safety model, system card numbers, comparison table) is in my Dots (AI) note.

That's it for today! ✨

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe