OpenAI launched Dots on September 29, 2026, as persistent personal agents powered by GPT-6 Astra. They work through a cloud computer and connected applications. For adopters, the important question is which background work is allowed and which actions require review.
Availability is a phased rollout
The announcement starts rollout on Pro and Business Premium in eligible markets. Enterprise workspaces, including Edu and Healthcare, can enable a beta through an administrator. Specialist Dots with organizational identities are described as focused enterprise pilots.
A business should distinguish a person delegating their own work from an agent owning a company process. The second case needs an identity, access lifecycle, escalation owner, and measurable service expectations. A personal-agent trial does not establish that these enterprise arrangements are already configured.
Background research and delegated actions are different
OpenAI says proactive background research uses connected-app tools restricted to reading. Dots can also perform delegated tasks with action review and rules. The separate safety and privacy explanation describes the safeguards surrounding that behavior.
This distinction is useful when setting expectations. Reading an inbox to prepare a briefing is different from sending replies or changing records. Write the boundary in concrete terms: which sources may be read, which output should be prepared, who reviews it, and which action must wait. Do not use a broad goal such as “manage my accounts” as the only operating instruction.
Persistent memory needs a correction path
A recurring agent can accumulate preferences and context that improve continuity. It can also retain an outdated assumption. If a customer changes requirements or a project owner leaves, there needs to be an obvious way to correct the agent’s understanding and inspect the work that depended on it.
Begin with a repeatable, reviewable artifact such as a morning project brief. Compare its references with the original sources and note repeated omissions. Review output over several runs, including quiet days when the correct behavior may be to do nothing.
Included access is not unlimited downstream work
OpenAI distinguishes conversations with a Dot from tasks it starts or manages in Codex and ChatGPT Work, which count toward those products’ usage limits. Treat these as separate measurements when deciding how much recurring work to delegate.
Our recommendation is to use a narrow task with read access first, then add write capabilities only when the team can explain the approval and recovery path. Dots can still make mistakes. Consequential work should remain inspectable even when the agent runs while its owner is away.