Here’s how most people will start with Dots: open ChatGPT, spin one up, give it five goals and connect every tool they have. Then they’ll wonder why it went sideways on day three.
Don’t do that.
An always-on agent is closer to a new hire than a new app. You wouldn’t hand a new hire every password and ten priorities on their first morning. Same here. This is the order I’d follow to use ChatGPT Dots well, as of October 2026.
Before you start: three checks
- Availability. Dots are available in ChatGPT on certain paid plans in eligible markets. Check OpenAI’s current plan page for your account.
- Policy. Decide what data a Dot may and may not touch. If your company doesn’t have an AI policy yet, use my AI policy outline as a starting point.
- The job. Pick one goal you can describe in a paragraph. If you can’t explain it to a new hire, you can’t explain it to an agent.
Step 1: Name your primary Dot
Every setup starts with a primary Dot you can name. Name it after the job, not after a personality. “Weekly Pipeline” tells you what it’s for six months from now. A cute name doesn’t.
You can launch it from ChatGPT or from Codex. Follow OpenAI’s current setup flow for your account — the screens will change faster than this page.
Step 2: Give it one goal, written like a brief
Dots pursue goals continuously in the background. That’s powerful, which is exactly why the goal has to be tight. Use this brief:
Goal: [WHAT OUTCOME YOU WANT, IN ONE SENTENCE] Why it matters: [WHO USES THE RESULT AND FOR WHAT] Allowed tools and sources: [ONLY THE ONES THIS JOB NEEDS] Done looks like: [THE EXACT DELIVERABLE AND WHERE IT GOES] Ask me before: [SENDING, SPENDING, DELETING, CONTACTING ANYONE NEW] Report back: [WHEN AND IN WHAT FORMAT]
Good first goals are internal and low-risk: a weekly summary of a project channel, a running research list, a follow-up queue you approve before anything goes out.
Bad first goals: anything that talks to customers, moves money or deletes files.
Step 3: Connect only what the job needs
Individual Dots can be given their own identities, credentials and tools. Use that to your advantage. Give this Dot access to exactly what its goal requires and nothing else. Read-only beats read-and-write until you’ve seen how it behaves.
If you’re connecting business tools, it helps to understand how AI plugs into systems in the first place. My explainer on the Model Context Protocol covers the basics and the security questions to ask.
Step 4: Decide where you’ll talk to it
You can message Dots through Slack, Microsoft Teams and other workplace platforms, and OpenAI has announced text messaging as coming. Pick one channel and stick to it. An agent that reports in the place you already check gets reviewed. One that reports somewhere you forget gets ignored.
Your first-week training plan
Day 1 — Brief and baseline. Give your Dot the goal brief. Ask it to restate the goal, the limits and what it plans to do first. Fix any misunderstanding now.
Day 2 — Watch the first output. Read everything it produced. Note what was right, what was wrong and what was missing. Tell it — specifically.
Day 3 — Tighten the instructions. Turn yesterday’s corrections into the brief. “Use the client’s first name” beats “be more personal.” Generic in, generic out.
Day 4 — Test an edge case. Give it a messy input on purpose. Does it ask, or does it guess? You want an agent that asks.
Day 5 — Check its reach. Review what it accessed. Anything it didn’t need? Remove it.
Day 6 — Let it run longer. Shift from watching every step to reviewing a daily summary.
Day 7 — Scoreboard review. Did it finish? How accurate was it? How much did you fix? Could you see what it did? If those answers are solid, it’s ready for a second goal — or a specialist.
Step 5: Add a specialist Dot (only when ready)
OpenAI describes specialist Dots and teams of Dots working together. When your primary Dot is reliable, a specialist makes sense for a distinct job with distinct tools — research, reporting, drafting.
Give the specialist its own identity and credentials, its own brief and its own owner on your team. Then decide how work hands off between them. I explain handoffs and accountability on multi-agent systems, and the step-by-step build logic applies to any platform on how to build an AI agent.
What never to give a Dot (or any agent) unsupervised
- Messages to customers that haven’t been reviewed by a person.
- Payments, refunds or anything that moves money.
- Deletions of files, records or accounts.
- Legal, HR, hiring, lending or housing decisions — those carry real discrimination and compliance risk.
- Calling or texting customers. AI voices fall under TCPA consent rules after the FCC’s February 2024 ruling, business texting needs consent and A2P 10DLC registration, and Do Not Call applies. Get counsel first.
Let AI do the work. You do the thinking.
Where Dots fit next to my picks
For my own setup, I pick Grok Business for business, Muse for personal life and Claude for building custom agents — my reasons are on the best AI agents hub. Dots are a strong fit when your team already lives in ChatGPT, Slack or Teams. If that’s you, this first-week plan is the right way in. For the bigger picture, go back to the OpenAI Dots overview.
Your next step
Write the one-paragraph goal brief today, before you open ChatGPT. If you’d rather have someone set up an agent around your real processes with you, see how Chief works. And if you want to watch agent setups like this happen on real work, I post them on YouTube.
Questions people ask
How do I start a Dot?
As of October 2026, Dots launch from ChatGPT or Codex on certain paid plans in eligible markets. Follow OpenAI’s current setup instructions for your account, then name your primary Dot and give it one goal.
When should I add a specialist Dot?
Only after your primary Dot handles its first goal reliably for about a week of daily check-ins. Adding agents to an unreliable setup multiplies the mess.