OpenAI launches Dots: define the tasks and permissions of persistent agents

According to OpenAI, Dots can use its cloud computer, interact with applications, and handle multiple tasks. The announcement mentions a rollout for Pro, Business Premium, and Enterprise offerings in eligible markets, as well as a preview of specialized agents for organizations.

The examples in the press release illustrate reported capabilities and uses. They do not prove that all processes in your company can be executed with the same level of reliability. Daillac did not test Dots for this article and does not confirm the eligibility of any specific subscription.

"Helps sales" gives a very broad objective. "Each morning, identify requests received without a responsible person and prepare a proposed assignment" defines a task, a source, and an output. The agent can start in read-only mode and submit a result for validation.

Also specify stopping conditions: inaccessible source, ambiguous result, unusual volume, or sensitive information. A permanent agent should not continue improvising just because it has not received new instructions. It must be able to report that a decision or permission is missing.

This separation makes the trial easier to evaluate. A team can measure the quality of the proposal before delegating the action. The agent validation framework complements this approach when rules must be met before an operation.

Inventory connected applications, accessible folders, and stored data. A person may have rights that exceed the needs of the mission. The agent should not receive these rights simply because they work for this person.

Define how to correct a preference or saved information, how to review actions, and how to remove access. Continuous automation requires a responsible person available for exceptions and changes. Authorization controls remain relevant even when an action is requested in natural language.

Track accepted tasks, corrections, interventions, and costs. Keep examples of failures, not just successes. An agent who prepares ten summaries of which eight need to be redone may shift the work without reducing it.

A AI automation should also be evaluated when the data changes, when a connection times out, and when the manager is away. These situations are part of the operation, not just a model test.

A permanent agent does not remove operational responsibility nor the need for validation. The announcement does not guarantee performance for a sector, error-free autonomy, or universal access to the product. A limited initial mission, a documented trial, and verified revocation provide a stronger basis than general deployment.

Fictional mission: prepare each morning the list of business files to review. The agent can read files within their scope and draft an internal note; they cannot modify a business step, send an email, or delete a file. A proposed modification goes to the responsible person. Acceptance testing checks the note and the permissions actually used. This matrix serves to make the separation between reading, proposing, and executing concrete.

OpenAI — Introducing Dots · OWASP — API Security Top 10, 2023 edition. Sources consulted on 1 October 2026; the grids and sample decisions are an editorial analysis by Daillac.