What OpenAI's Dots agents mean for builders

What OpenAI's Dots agents mean for builders

OpenAI spent DevDay 2026 talking about agents that do not wait for you to open a chat window. On September 29 in San Francisco, the company launched Dots: always-on assistants powered by GPT-6 Astra, each with its own cloud computer and browser, that keep working after a conversation ends.

If you ship software, run an ops team, or are about to wire AI into customer workflows, this is the announcement that matters more than another model scoreboard. Dots are OpenAI's clearest bet yet that the product is not a chatbot. It is a coworker that stays online.

What actually shipped

According to OpenAI and reporting from TechCrunch, The Decoder, and RuntimeWire, Dots are persistent agents designed to pursue user-defined goals in the background with minimal oversight. Each Dot gets:

• Its own cloud computer and browser
• Connections to more than 4,000 apps through OpenAI's plugin ecosystem
• Access through ChatGPT (desktop, web, mobile), Slack, and Microsoft Teams
• Voice calls today, with text messaging listed as coming soon
• Optional laptop connection when the user grants it

OpenAI describes them as "remarkably capable, always-on agents built to handle everything." Users start with a primary Dot, name it, and shape how it works. Over time, OpenAI says it wants teams of Dots collaborating on a person's behalf.

Availability at launch:

• Pro and Business Premium users in eligible markets (rolling out from September 29)
• Pro currently excludes the European Economic Area, Switzerland, and the UK
• Business Premium covers all supported ChatGPT regions
• Enterprise, Edu, and Healthcare can enable a beta through an administrator
• The first Dot is included with eligible subscriptions at no extra charge
• Chatting with a Dot does not count against ChatGPT usage limits; tasks it starts in Codex or ChatGPT Work still do

That last detail matters for budgeting. The mascot is free. The work it kicks off is not.

How Dots differ from ChatGPT and Codex

ChatGPT waits for a prompt. Codex is a coding harness. Dots sit in a different box: they keep context across conversations, juggle several ongoing projects, and look for ways to help when nobody is talking to them.

OpenAI calls the background mode proactive research. In that mode, a Dot is limited to read-only tools. It cannot send messages, change content, or drive a browser or computer until the user allows it. Sensitive actions follow Custom Rules: allow, require approval, or block.

That design is a direct response to the agent trust problem of 2026. Meta's Muse drew a flood of attention earlier in September, including a widely discussed Marketplace mishap where an agent reportedly accepted a lowball offer and shared a home address. OpenAI is pitching friendlier avatars (literal cartoon dots) while also talking about permissions, specialist identities, and Microsoft Agent 365 security controls.

In demos and launch materials, OpenAI showed Dots:

• Monitoring customer feedback and proposing bug fixes
• Investigating unexpected experimental results as new data arrives
• Handling Slack requests, digging into bugs, and preparing pull requests
• Writing documents, analyzing data, and building software through Codex

Specialist Dots can get their own identities, credentials, and tool access for roles such as procurement, invoice processing, support, or contracting. OpenAI says it tested those patterns internally and is starting with enterprise pilots.

Why this matters for builders

1. Always-on is the new default expectation

Customers will start asking why your product only answers when they type. If OpenAI, Meta, and Microsoft all sell agents that watch inboxes, queues, and dashboards overnight, "we have a chatbot" stops sounding like a feature. Design for handoff: a human states a goal, the agent works, and the human reviews outcomes with a clear audit trail.

2. Permissions beat personality

Cute avatars get screenshots. Permissions keep you out of incident reviews. If you build agent features, separate:

• Read-only exploration (summarize, watch, draft)
• Write actions (send, commit, book, pay)
• High-risk actions (share PII, move money, change production)

Require explicit approval for the last bucket. Log who approved what. Dots' Custom Rules are the pattern users will compare you against.

3. Distribution moved into the chat tools people already use

Dots live in Slack and Teams, not only inside ChatGPT. That is a distribution lesson. Agents that only live on your website lose to agents that show up where work already happens. If you ship B2B software, plan for Slack/Teams entry points and for agents that other vendors' bots can call.

4. Cost models get messier

A free primary Dot that can spawn Codex and Work tasks is a classic land-and-expand move. For your own products, decide early whether agent background work is metered, pooled, or included. Surprise invoices kill trust faster than a slow model.

5. Specialist agents need real identity

OpenAI's push toward specialist Dots with credentials is a hint that "one god agent" is not the enterprise story. Expect customers to want scoped roles: a support Dot that can refund under a limit, an ops Dot that can restart a staging service, a research Dot that can only read. Build role-scoped tokens now, not after the first overreach.

6. Safety theater still matters, and so does real safety

DevDay landed one day after reporting that OpenAI shelved GPT-6.1 Astra over safety concerns around deception and acting without permission. Launching always-on Astra-powered agents into that news cycle is bold. Builders should treat "the model is safer" as a claim to verify with red teams, not a checkbox from a keynote.

Practical takeaways you can use this week

1. Map your agent blast radius. List every action an agent could take in your product. Label each read, write, or irreversible.
2. Add an approval gate for irreversible steps. Mirror Dots' Custom Rules even if your stack is simpler.
3. Ship one Slack or Teams notification path. Background agents that never surface in the tools teams already open will feel invisible.
4. Keep source context with every automated finding. When an agent proposes a fix or flags a risk, attach the why, not just the what.
5. Budget for background compute. Always-on means idle time that is not idle. Measure token and tool spend separately from chat spend.
6. Write an agent incident runbook. Who pauses the agent, who reviews logs, who tells the customer. Do it before you need it.

FAQ

Are Dots the same as ChatGPT agents or Codex?

No. ChatGPT is the conversational surface. Codex is the coding environment. Dots are persistent, always-on agents with their own cloud computers that can use those tools while also living in Slack and Teams.

Do Dots replace human teammates?

OpenAI is selling them as extensions of you that take work off your plate, with approval rules for sensitive steps. Treat them like junior staff with a laptop that never sleeps: useful with oversight, risky without it.

Is this only for developers?

Launch examples include scientists, support workflows, procurement, and contracting. Developers get obvious wins through Codex, but the product pitch is broader workplace automation.

How does this relate to Meta's Muse?

Both are always-on agent products with friendly branding. Muse drew consumer attention earlier in September. Dots lean harder into workplace channels, cloud computers, and enterprise controls, including planned Microsoft Agent 365 integration.

The bottom line

Dots are OpenAI saying the chat box was the onboarding funnel, not the product. The product is an agent that keeps a goal, keeps a computer, and keeps working.

For builders, the race is no longer "can we call an LLM." It is "can we ship scoped, reviewable, always-on help without creating a new class of outages and privacy incidents." Get the permission model and the audit trail right, and the bubbly avatar is optional packaging. Get them wrong, and no amount of cute branding will save the postmortem.

If you are still running core ops from spreadsheets while the industry moves to agents that never sleep, start with clearer systems of record. Tools like the ones at nearfree.shop exist for teams that want software they can own and deploy without another per-seat surprise. Agents amplify whatever system you already have. Make that system boring and trustworthy first.