I Gave OpenAI's Dots Real Jobs for a Day
I set up OpenAI's always-on agent and handed it my shopping, a Montreal trip and my AI agents. Here is what worked, what broke and whether it's worth Pro.
I set up OpenAI's new Dots agent, named it László, and handed it real jobs for a day: my Amazon shopping list, a trip to Montreal, and the project board my AI agents work from. The shopping failed and the trip worked, but what impressed me most was a billing problem it caught in my email without being asked. Below is what it did, where it broke, and whether it's worth paying for.
What a dot is
A dot is an agent that lives inside ChatGPT, runs on GPT-6 Astra, and keeps working in the background around the clock. It has its own cloud computer and its own browser, and you can let it use your computer too. You message it, call it, or talk to it in Slack or Teams. You get one dot, and it comes with the Pro plan, which starts at $100 a month.
ChatGPT now has four things that sound alike: ChatGPT, ChatGPT Work, Codex and Dots. The easiest way to see the difference is one job. Say I have a YouTube video coming out in a couple of days.
- ChatGPT gives me title ideas and some script advice, and it's only good if a project already knows my voice.
- ChatGPT Work reads my past scripts in Google Drive, researches YouTube in its own browser and writes a research doc.
- Codex pulls viewer retention and comments from the YouTube Data API and uses them to shape the outline.
In all three, I hand over each step, check it and give feedback. With a dot, I say "help me get this video out in the next couple of days," and once it knows my process it works out the steps itself. It notices the thumbnail isn't done, kicks off a build in Codex and watches it, has another agent write the blog post, and after launch messages me in Slack about what viewers are saying.
You could build this yourself with Work or Codex. I've tried. You need a cloud environment connected to GitHub, scheduled tasks, an agents.md, a memory file, some way for it to learn, and a pile of threads and projects. A dot does that out of the box.
How it compares to Grok Bot and Meta's Muse
- Grok Bot. You set up a separate bot for each job: a comment reader, a script writer, a thumbnail planner, a researcher. Each bot sticks to its one job, and you have to go to each of them separately. It runs on Grok 4.7 and starts at $20 a month, and people say $20 buys very little usage. It has X built in.
- Meta's Muse. A consumer app with Instagram and Facebook built in, for everyday stuff rather than work.
- Dots. One dot handles everything, on one of the best models available, and it already knows your ChatGPT history, which makes onboarding faster.
Setting it up
Your dot sits at the top left of ChatGPT. Setup is a few clicks: connect Gmail, Contacts and Drive, then choose where it can work. Mine picked its own name, László.
- Use the desktop app to connect your computer. The web version can't reach your Mac. The option only appeared after I restarted the desktop app, and I had to redo onboarding there.
- It works on one thing at a time. You can't start a new chat with it. Side jobs show up under Recent activity, a bit like chats inside a project.
- Talk to it in Slack. The mobile app had no dot the night I left my office (it showed up the next morning). I connected László to Slack and kept giving it jobs on my way home. Each job gets its own thread and the replies stay organised. I still prefer this to the mobile app and to the Codex app.
Its first job was to learn my YouTube workflow from the agent and skill files on my Mac. It found the repo, mapped every skill (thumbnails, hooks, scripts, blog posts, benchmark builds) and knew which videos were in progress. Codex could do that too, so it was a warm-up.
Job 1: Personal shopper
I gave it a shopping list and asked it to find everything on Amazon and buy it. Amazon blocked its browser straight away. When I told it to build a way around the block, it refused, and fell back to Chrome on my computer.
From Slack it kept me posted, sent a screenshot of the plates, asked "green or blue?", and added or removed items as I answered. That part felt good. The shopping itself is nothing Work or Codex can't already do in a browser, and plenty of admin sites will block its browser the same way. As a personal shopper, it's not there yet.
Job 2: Travel planner
A trip is one of those jobs ChatGPT half does: it gives you links and you still do the clicking. I asked László out loud:
So I've got this trip to Montreal scheduled October 15th to the 18th. I need an Airbnb and I need a flight. Please find an Airbnb that is centrally located, relatively inexpensive, close to the city center. If Airbnbs are way more expensive, you could also just do a hotel. I need you to find the cheapest flight that's a direct. I want an aisle seat. I'm bringing a carry-on. So go ahead and do all of that and then put together a plan of things to do. I want to, of course, walk Mont Royal, but is there anything else, just given what you know about me, that I would find interesting? And find a couple good restaurants that have like a nice, cool aesthetic.
It worked fast in its own browser, on Airbnb and then Air Canada, neither of which blocked it.
- For the Airbnb, it found a central, cheap place with good reviews and compared it with nearby hotels. When I asked about the 4.75-star listings, it found repeated complaints about a squeaky Murphy bed.
- For the flight, it found a direct Air Canada flight and filled in my details, using the email and phone number it found in my travel records.
- For the plan, Mont Royal's cross is under construction, so it built the walk around Beaver Lake instead. It added an architecture stop, an artist lecture on the Saturday, coffee and dinner.
It stops at checkout, so paying is still my job. As a travel planner, it's there.
Everything else I threw at it
A marathon training dashboard
I'm running the New York City Marathon in November with Zone 2 training from a Skool community. I asked László to learn the method from the community, check the last three months of pace and heart rate from Strava and Garmin, estimate my race, and message me after every run. Garmin's API needs a developer account, so it signed up for Intervals.icu on its own, pulled my runs and built a dashboard called Steady. The app is rough, with the same color palette Astra always reaches for, but it set up the whole data connection without me.
Managing my AI agents
My agents track client work in Code Queue, a Linear-style board I built with an MCP server. I almost never look at it. László found Code Queue on my Mac, checked the MCP connection, read the API and scheduled a status summary in Slack every weekday at 9 PM.
Then I called it. It said 137 issues were in progress and 122 were in review. Most of those were stale, so I asked it to propose which to close, send the list in Slack and close them once I approved. About a minute after I hung up, the Slack message arrived: "Code Queue cleanup." None of the other ChatGPT products do this.
Spanish practice calls
I asked it to call me every night at 8 PM to practice Spanish. It can't make outbound calls yet, so you call it from the dot app. It takes a moment to pick up, then holds a conversation with almost no delay.
Selling my dad's business
My dad runs a laser printer repair company and wants to sell it. László researched it and built a website holding the sale plan: a buyer shortlist, what transfers, deal and price, the sales process, and the first 30 days. It also made a spreadsheet of 136 potential buyers with a fit score for each. Then I asked it to check in with him once a week and report back to me. He's more likely to talk it through with an AI checking in on him than to bring it to me.
Growing my newsletter
It didn't know about my newsletter from my ChatGPT history, so I filled it in and asked it to study how The Hustle, The Rundown and Morning Brew grew. It wrote a 30-page PDF. The first summary was generic; when I pushed for specific tactics, several were good. It's now keeping the goal in memory and sending me ideas.
The thing I never asked for
I woke up to a message: Vercel's third attempt to collect $22 had failed, with another retry on October 2nd. I never asked it to watch my invoices. I had only connected Gmail. It also noticed I had already used $15 of my $20 monthly Vercel usage credit, which I didn't know, because I filter those emails out. Since it can see my repos, I can now ask it how to bring that usage down.
What other people are doing with Dots
Most of what I've seen from other users is admin and travel:
- One dot read a cancellation email, checked live tracking on an earlier train, found it running late and got its owner on it.
- Another noticed an Uber was booked to a different terminal from the arriving flight.
- One went through a live chat to cancel an in-flight wifi subscription and got the refund.
- One spotted half an hour between two events, asked how its owner planned to get there and suggested cheap parking.
- One kept emailing a hotel for two days, then found the general manager's name and emailed them directly.
OpenAI's own demos were a board deck and a website. A dot does that kind of build better by handing it to another Astra agent that already has your skills and deploy setup, and acting as the project manager.
Is it worth it?
- It's always on. It keeps its own computer, so it picks up right where it left off in its own browser, with no laptop open and nothing spinning up each time.
- It watches your stuff. The more apps you connect, the more it digs through them. If your work lives in Outlook, Notion, monday.com or a HubSpot CRM, connect those. Based on what it did with my Gmail, I expect the same there.
- You're still the one juggling. It brings you approvals, decisions and more messages. That's one more worker to manage, however productive it is.
- Context rot. I haven't hit it yet, but I already have a lot of sessions with one dot. Astra has a million-token window, so eventually it compacts and forgets. I want to see how it handles memory and learning, which you used to manage yourself with tools like Hermes or OpenClaw.
If you're already paying for Pro at $100 or $200 a month, turn it on, connect your plugins and talk to it about your day. It will start working. At first I wasn't convinced it was a step up from ChatGPT Work or Codex. After a full day with it, I see the potential.
Comments
Sign in to join the conversation.