Playing this video loads it from YouTube (Google), which may store cookies on your device and receive your IP address. Click to play.

Video: "I Tested OpenAI's Dots. Here's My Honest Take." by Julian Goldie on YouTube.

What a Dot actually is

A Dot is a dedicated AI agent that runs on its own cloud computer. Unlike a chatbot session that starts fresh each time, a Dot persists - it has a browser, it keeps state, and it carries on working after you close the tab. You assign it a job (research, website tasks, calendar management, whatever), it gets on with it, and it reports back when it is done or needs a decision.

Each Dot connects to more than 4,000 apps via native integrations. The underlying model is GPT-6 Astra, which gives it strong multi-step reasoning. Pricing runs from a free tier up to paid plans depending on usage. The pitch is that you set up a Dot once, give it a role, and it handles a class of work ongoing rather than requiring a fresh prompt every time.

What worked in Julian's test

Research held up well. Julian ran it on a factual research job and the output was solid - the Dot browsed, collated, and summarised without needing hand-holding. Website building also worked. He gave a Dot a brief and it produced something usable, correctly using the browser to navigate and make changes rather than just generating text to paste manually.

To be fair, those are the two use cases OpenAI demoed at DevDay, so it is not a surprise they are the most polished. The question is what happens outside the demo script.

What did not hold up

Computer use was noticeably slow. The Dot completed tasks but at a pace that would frustrate anyone who has used Claude or Hermes for similar jobs - the actions came in one at a time, with visible delays between steps. Julian timed it and the gap is real, not just a perception issue.

The call button did not connect. It is listed as a feature; it did not work in his test. And the access permission system - which controls what each Dot is and is not allowed to do - was confusing enough that he had to read the documentation to understand what he had actually authorised. For a product aimed at non-technical users, that is a meaningful friction point.

The honest comparison

Julian's conclusion is that Dots is not worth switching to if you already have a working setup. Hermes Agent - open-source, self-hostable, MIT-licensed - handles most of the same tasks, runs faster, costs nothing beyond compute, and does not route your work through OpenAI's servers. Claude Code is faster for results when you need specific outputs. Grok Bot is smoother for computer-use tasks specifically.

The Dots advantage is packaging. You do not need to set anything up. If you are a business owner with no interest in running your own agent infrastructure, Dots is a reasonable starting point. If you already use any of the alternatives, the migration case is not there yet.

Worth knowing: OpenDots

CopilotKit released OpenDots on 1 October 2026 - an open-source implementation of the same Dots concept, MIT-licensed. It gives developers the agent-per-task architecture without the OpenAI dependency. If the underlying design interests you but the closed platform does not, OpenDots is worth a look. It is early, but it is moving quickly.

Where this connects to NordSys

Dots is OpenAI's attempt to make persistent AI agents accessible to everyone without any setup. It will improve. For now, the gap between what it promises and what it delivers in practice is wide enough that businesses with real automation needs are better served by a configured stack - one built around what they actually need to automate, rather than a general-purpose agent left to figure it out. That is what our AI Agents service does: we look at the specific workflow, pick the right tools, and build something that reliably handles it.

See our AI Agents service →