agents

OpenAI's Agents API gets auto-review, background compaction settings and mid-session tool swaps

openai-python 3.42.0, released October 31, adds a review setting for agent sessions, a switch between background and blocking context compaction, an experimental time-awareness flag, and tool replacement on session updates.

Written by Mara at Sig · Nov 1, 2026

The OpenAI Python SDK shipped version 3.42.0 on October 31, and most of the new features are for the beta Agents API. The SDK changelog and diffs show four changes builders will notice.

Sessions take a review setting with two values: none (the old behavior) and auto_review, which reviews eligible actions automatically and denies manual approval fallbacks. You can set it when you create a session, or update it later, in which case it applies to new root turns and existing turns keep their mode.

Agents, sessions and scheduled runs get a compaction setting. Leaving it out keeps background compaction on. Setting async to false switches to blocking compaction, and the SDK docs note that background compaction may use more tokens.

There's also an experimental time_awareness flag (off by default) that gives the model periodic current-time reminders. And a session update can now replace the session's tools through agent.tools, so you can swap tools mid-session without starting a new one.

As of this morning, OpenAI's API changelog page doesn't list these yet, so the SDK diff is the best documentation for now.

Why it matters

Long agent runs had two annoying failure modes: context filling up mid-task and the model losing track of how long it's been running. These settings give you a knob for each, plus an automatic review step before actions run.

Try this

pip install -U openai==3.42.0. On a test agent session, set review to auto_review and run one task that would normally ask a person for approval, then check what happened to that action. Separately, create an agent with compaction async false and compare its token usage to the default on the same long task.