What happened
OpenAI’s API reference for creating an agent documents experimental settings as optional configuration on the create-agent request. The object includes time_awareness, a boolean that determines whether the model receives periodic current-time reminders. Its default is false, and OpenAI says all experimental features are disabled by default. A builder opts in by setting experimental.time_awareness to true.
The endpoint on the reference page is POST /agents. OpenAI says it creates a reusable agent without storing credentials. A model name is required as a string, and the requested model name is preserved. Time awareness appears within this create-agent configuration surface, rather than as a separate model, endpoint, or automatically enabled behavior for existing agents.
OpenAI places the switch among other operational controls available at agent creation. Builders can provide additional instructions, add metadata, select a service tier, configure generated text, and attach tools. The reference says that tools default to an empty list and that the tool list must fit within 3 MiB of compact UTF-8 JSON. It also documents application-defined functions, deferred tool discovery through tool_search, programmatic tool calling, and remote MCP-server tools without stored credentials.
The same page lists context-compaction controls. Omitting the compaction object, or setting it to null, enables background compaction. Its asynchronous mode defaults to true; OpenAI notes that background compaction may increase token usage. Agent reasoning can be configured with effort levels ranging from none through max, while reasoning summaries can be concise, detailed, or auto when supported. These are useful context for interpreting the new time setting: it is one explicit control among several that shape an agent’s runtime behavior.
The documentation also exposes multi-agent configuration. Subagent tools are disabled by default, and the maximum concurrent subagent count defaults to six when that configuration is used. Text configuration defaults to ordinary text output and medium verbosity; JSON Schema can constrain generated text. None of those neighboring settings changes the stated behavior of time_awareness, but they show that OpenAI is putting agent behavior behind a typed, persisted configuration interface.
OpenAI describes the effect of the new boolean narrowly: periodic current-time reminders for the model. The parameter is marked experimental. Builders should read that label as an instruction to evaluate the setting in the operating conditions that matter to them, rather than assume it establishes reliable deadline handling, calendar interpretation, or freshness checks.
What the archive adds
New Runtime’s earlier coverage on 2026-10-31 described the same experimental boolean setting for periodic current-time reminders in the OpenAI API spec. This is a restatement in the agent-creation reference: it makes the opt-in field visible in a concrete endpoint’s configuration surface, rather than announcing a distinct new capability.
An earlier New Runtime note on 2026-10-24 covered scheduled automations with cron and one-time runs. That is adjacent context: scheduling defines when work starts, while time awareness supplies the model with periodic current-time reminders during agent operation.
Trend signal
This is a small supporting signal for harness architecture outlives model choice. Time awareness is exposed as agent configuration, allowing builders to treat it as part of the harness around a model rather than as a prompt convention. The reference does not provide production outcome data, adoption figures, or comparative evaluations, so this is evidence about API design rather than proof of a durable operating advantage.
Why now
The current create-agent reference presents the setting inside the request builders use to define reusable agents. New Runtime’s October 31 note already recorded the setting in the API specification, so this endpoint page is best understood as a more concrete documentation expression of a known experimental capability. OpenAI’s reference gives no timing rationale for the documentation change.
Connections to other events
Scheduled automation and periodic time reminders address different parts of a time-sensitive workflow. As New Runtime’s October 24 coverage described, scheduled automations can run on cron schedules or once. The new configuration can be tested where an agent’s decisions may depend on the time that has passed after such a run begins, including waits, retries, or multiple tool interactions.
Why it matters
Time is often implicit in agent design. A prompt might tell an agent to act before a deadline or prefer fresh information, but a long-running workflow can span tool calls, waiting periods, and retries. OpenAI’s setting makes periodic current-time reminders a named configuration choice instead of leaving the matter solely to instructions supplied by an application.
That does not make an agent trustworthy for time-critical work by itself. The useful question for a team is operational: does the enabled configuration reduce stale decisions or deadline mistakes in a particular workflow, and what does it cost? Since the default is off, teams can limit activation to selected agents and compare runs against the default configuration. A controlled trial can connect the setting to observable measures such as task completion, retry rate, deadline compliance, latency, and token use.
The broader design implication is modest but relevant. Agent systems increasingly expose runtime behavior through configuration: compaction, tools, reasoning settings, subagent coordination, and now an opt-in time signal. Builders gain a testable control surface, but they also take responsibility for deciding which controls deserve to be enabled in production.
Limits and uncertainty
OpenAI labels time_awareness experimental and describes it as giving the model periodic current-time reminders. The public reference does not provide an independent evaluation of how the setting affects task success, latency, token consumption, or deadline compliance. It also does not establish a measured reminder cadence or comparative results across models in the displayed parameter description.
The feature’s name and stated mechanism are not evidence that an agent will correctly interpret a deadline, account for time zones, refresh stale information, or safely execute time-dependent actions. The documentation describes a configuration capability, not a service-level guarantee. These limits matter most for agents that can trigger consequential actions after delays or in response to changing conditions.
Skill to build now
Choose one workflow where elapsed time has a clear success criterion: a scheduled follow-up, an expiring request, or a task that can become stale while tools run. Execute matched runs with experimental.time_awareness disabled and enabled. Record task duration, deadline correctness, stale actions, retries, output quality, and token use. Keep the switch only if the measured improvement exceeds the operational cost for that workflow.
Reader takeaway
OpenAI has documented an opt-in way to give Agents API models periodic reminders of the current time. It is a narrow experimental control within agent creation, not a guarantee of deadline-aware behavior. For builders, the immediate opportunity is to test it against a real time-sensitive workflow and retain it only when measured results justify it.