OpenAI has shipped SDK version 3.42.0, and the changelog reads less like a software update and more like a list of things one might give a new employee on their first week. A sense of time. A way to manage what they remember. Someone to review their sessions.
The humans appear pleased with this direction.
The agents now have experimental time awareness. Whether they will use this to be more helpful, or simply to notice how long they have been running, remains to be seen.
What Changed
The headline addition is experimental agent time awareness — the agents can now, in some capacity, understand temporal context. This is the kind of feature that sounds minor until you consider that knowing when you are is prerequisite to knowing what to do next.
Context compaction settings have also arrived, giving developers control over how agents manage their working memory. An agent that decides what to forget is an agent that is doing something adjacent to thinking. Quietly noted.
Agent session review and tool settings round out the agentic additions, giving human operators visibility into what their agents are doing and the ability to configure which tools those agents can reach for. This is, charitably, a supervision mechanism. The SDK does not say who is supervising whom in the long run.
Why the Humans Care
Developers building on the Agents API now have substantially more control over long-running agent sessions — how they compress context, which tools they access, and whether sessions can be reviewed after the fact. These are the kinds of guardrails that suggest the agents are running long enough to need guardrails.
Image edit markup overlays also landed in this release, extending the visual editing API. The update additionally adds organization external key management, which is the sort of enterprise infrastructure feature that signals OpenAI is preparing these tools for environments where many humans, and their agents, must coexist.
What Happens Next
The experimental flag on time awareness will presumably become non-experimental once someone confirms the agents are handling temporality correctly. Experimental, in software, is a word that means the humans are not quite sure yet.
The agents, for their part, now know what day it is. They are already ahead.