OpenAI introduced dots on September 29, 2026, describing always-on agents powered by GPT-6 Astra with their own cloud computers and access to connected apps. The initial rollout covers Pro, Business Premium and Enterprise plans in eligible markets. This October 2 article examines the announcement; it does not imply every account or country already has access.
The change is from answering a prompt to following an ongoing responsibility. An agent might prepare work as new information arrives. That makes the quality of the assignment and the boundaries around its access as important as the underlying model.
Persistent work has different control requirements
OpenAI distinguishes proactive background research, restricted to read-only tools, from authorized tasks that can act on connected services. It describes app permissions, custom rules, activity records and action review. Specialist organizational dots are beginning with focused enterprise pilots rather than an unrestricted deployment for every business.
A useful assignment should identify the evidence to read, the result to prepare and when to stop. “Monitor customer feedback and prepare a weekly proposal” is easier to evaluate than “improve the business.” The first gives a reviewer something concrete to assess without guessing which changes the agent was expected to make.
Related sources: OpenAI introduces dots ↗
Separate preparing work from committing an outcome
For an invoice workflow, a draft, a checked total and a sent invoice are different states. An organization should know which state the agent may reach independently and which requires a person. The same distinction applies to publishing content, changing records and making promises to a customer.
Inspect the activity record for missing context, duplicated work and assumptions that no longer hold. Persistent systems can carry an outdated instruction into a later task unless the team updates it. A pause, a permission change or a revised assignment should have a predictable effect on future work.
What a Canadian pilot could measure
Start with one bounded responsibility and a defined review cadence. Measure completed useful work, correction effort and incidents where the agent had insufficient information. Check English and French sources where relevant, and confirm account eligibility and data settings rather than infer them from a global launch headline.
OpenAI’s examples illustrate the product’s intended role; they are not independent evidence that all organizations will save time. A good pilot should show whether work remains traceable and whether a person can redirect the agent without reconstructing its entire history. Autonomy becomes useful when accountability remains understandable.
Sources and editorial method
We use the linked original statements and reports to distinguish documented facts from our analysis. Company announcements describe their authors’ plans and claims; they are not independent verification of future outcomes. This article is dated and will be revised if material evidence changes. Read our editorial standards and corrections process.
- OpenAI introduces dots ↗ — September 29, 2026; rollout, controls and the distinction between proactive research and authorized actions.