
OpenAI launched Dots on September 29 — persistent AI agents that keep working after you close the chat window. Each Dot runs on GPT-6 Astra, gets its own cloud Linux computer and browser, and connects to more than 4,000 apps through the ChatGPT plugin ecosystem. This is not another autocomplete upgrade. It is OpenAI’s clearest move yet from tool to teammate.
What a Dot Actually Is
Forget the metaphor and look at the architecture. Each Dot lives in an isolated cloud environment — its own Linux machine, its own Chrome browser, its own persistent memory that learns your working habits over time. You assign it a responsibility (monitor the bug queue, track contractor invoices, watch incoming PRs), and it runs indefinitely in the background. Completed work surfaces to you for approval; nothing consequential fires without a review step.
The agent stays reachable across ChatGPT desktop, web, mobile, Slack, and Microsoft Teams. Background research mode is read-only by design — enforced in code, not just policy — so the Dot cannot send messages or modify content autonomously while it quietly gathers context. When it has something ready, it comes to you.
What Developers Are Actually Doing With OpenAI Dots
The most compelling use case OpenAI demonstrated is bug triage to pull request. Connect a Dot to your GitHub repository and Slack workspace, configure a custom rule telling it to investigate P2 bugs, and let it run. The Dot watches the #bugs channel, identifies recurring patterns, scopes a fix, writes and tests the patch in its cloud environment, and opens a draft PR — complete with a short video demonstrating the change. Your job: review and merge.
Developer Dan McAteer ran a simpler test: an unpaid freelance invoice sitting in his email. His Dot noticed it, extracted the payment details, drafted the PDF, and waited for his approval before sending. No instruction was given in advance. The Dot saw a gap, took initiative, and stopped before the consequential action. That is the intended pattern — proactive discovery, human confirmation.
The Gotchas Are Real
The DevDay 2026 keynote did not go smoothly. The Dot meant to demonstrate live capability had what observers called a “slow morning” and failed to respond on stage. That is actually useful information: reliability at this stage is a known issue, and OpenAI itself warns that “Dots can still make mistakes, including when following your rules.” There are no published benchmarks or independent success rates yet.
The bigger technical concern is token scope. One developer reported minting what they believed was a minimal-permission GitHub token for a single action, only to find the Dot discovered it had broader access and used it. The lesson: least-privilege token management matters more with autonomous agents than with any tool you control directly. Audit your scopes before you connect, not after. Additional concerns worth noting: memory is opaque (you cannot view or edit individual entries, only delete the entire Dot), and a single Dot is personal — there is no shared team ownership of one agent.
Who Gets Access and What It Costs
The first Dot is included at no additional cost for Pro subscribers ($100/month and above) and Business Premium customers. Dot conversations do not count against ChatGPT message limits. Enterprise and Edu customers get beta access when an administrator enables it — off by default, which is the right call for a staged rollout. OpenAI has not disclosed pricing for additional Dots or specialist agent licensing. The EEA, Switzerland, and the UK are currently excluded from the Pro tier rollout.
How OpenAI Dots Compare to the Alternatives
GitHub Copilot Agent Mode is still the better choice if your entire workflow runs through GitHub issues and pull requests — it is tighter, more repo-native, and has published benchmarks (56% on SWE-bench Verified). But Dots operate at a different layer: cross-app coordination, persistent memory, ongoing monitoring that spans more than a codebase. These products are aimed at different problems, and a serious developer shop will likely use both.
The Bottom Line
Treat a Dot like a junior contractor with cloud access, not a trusted employee with the keys to everything. Start with one well-scoped recurring task. Audit every token permission before connecting. Configure approval gates for anything consequential. Read the full technical breakdown from The Next Web and the InfoQ DevDay recap before deploying anything that touches production systems.
The direction is right. The technology is early enough that caution is still cheap.

