OpenAI’s Agents API documents Scheduled Automations: save a schedule that sends work to an agent session. Timing can be `periodic` with a five-field cron `expression` plus a named `timezone` (for example `America/New_York` or `UTC`), or `once` with `run_at` as a future Unix timestamp in seconds. The target can be `new_session` (agent, environment, and initial `input`) or `existing_session` (session ID plus reminder `instructions`). Creating an automation saves it as `active`; it does not run the task immediately. A periodic automation’s first occurrence is the next matching time.
The reliability pitfall is the invocation status. Each run creates an invocation—a record of the attempt to deliver a task to its target session. Status starts as `queued` and becomes `completed` when delivery succeeds. For a new session, that means the session accepted its initial input; for an existing session, that it accepted the automation notification and the request to start work. Completed does not mean the agent finished. Follow the invocation’s `session_id` through session events and the turn’s outcome to see whether the work actually completed.
Timing and queuing matter for monitors and dashboards. The schedule is the intended clock time; the Agents API adds a small randomized delay on recurring runs to spread load. If the target session is busy, the task waits until the session is available. A scheduled run does not interrupt an active turn, and it does not supply a missing tool result or approval—your application must still handle required actions.
Pause or delete stops future scheduled (and, for delete, manual) invocations. Neither stops work a session has already accepted, and a delivery already in progress may still reach its session. To stop current work, cancel the active turn.
You can also enable the `automations` tool so an agent schedules follow-ups in its current session (create, list, update, pause, resume, delete). The API sets that session as the target; the agent cannot use the tool to open a new session or manage another session’s automations. The tool is off by default; your app can schedule work without enabling it.
The create path still sits under the Agents API beta surface (requests use the `OpenAI-Beta: agents=v1` header). It is aimed at developers with Agents API access. That is not the same as a ChatGPT subscription gaining a new free entitlement, and it is not the same product as ChatGPT Team Tasks—those are workspace-level ChatGPT automations, not Agents API session schedules. See also What ChatGPT Team Tasks Do and How Billing Works.
For calling create, listing invocations, or measuring end-to-end finish times on your own account, we have not tested yet; if you have, you can reach the site owner through the contact link in the footer. Details above are for reference only. Independent scheduling fees, and which accounts or regions have access, are not confirmed in the public docs we cite.