OpenAI Agents API Auto Review SDK 3.42 is the quietest important release of the week. On October 31, 2026, OpenAI shipped Python SDK version 3.42.0, adding compaction settings, session review, experimental time awareness and organization key management to the beta Agents API.
None of this is a new model. It is the harness — the loop, safety checks and memory management around the model — getting controls developers can finally set themselves.
What shipped, and what is on by default
The defaults show what OpenAI trusts and what it does not trust yet:
- Auto Review: off by default. Like Codex, it evaluates the potential for inappropriate behavior before executing commands. You must opt in.
- Background compaction: on by default. Long sessions get summarized in the background instead of blocking the agent while it compresses context. Set the flag to false for blocking compaction instead.
- Time awareness: experimental. The agent gains a sense of time passing, though the SDK notes do not specify whether that means calendar awareness, elapsed-time tracking, or both.
- Session review and tool settings. Developers can now inspect completed sessions programmatically, simplifying debugging and audit work, and update tools mid-session.
- Organization external key management. Tighter credential control, with callable API keys kept scoped to each request.
The settings landed in three SDK pull requests merged October 29 and October 31. The API changelog does not list them, so the SDK release notes are the precise record.
Why compaction settings matter more than they sound
Compaction is how a managed agent handles a long session: when the conversation outgrows the context window, the harness summarizes earlier parts so work continues. Until now, that happened server-side, invisibly.
Making it a client parameter changes the trade-off you control. Background compaction keeps the agent moving but may increase token usage, by an unstated amount. Blocking compaction pauses the agent while it summarizes: slower, but easier to reason about and audit.
For a long-running task — a multi-hour migration, a research sweep across dozens of sources — the wrong compaction choice quietly drops the constraint that mattered. A client-visible setting lets you test both modes on your own workload.
The beta warning still applies
The Agents API is still in public beta, launched September 10, 2026. Calls use a beta client path and header, and OpenAI has published no latency or cost-saving numbers for these settings.
That means two disciplines: do not build a client promise on beta behavior that can change, and keep your acceptance record — task, model, configuration, timestamps and result — outside the conversation, so a compaction change cannot silently rewrite what “done” meant.
The evidence: release facts
| Detail | What the source says |
|---|---|
| Released | October 31, 2026, Python SDK 3.42.0 via PyPI |
| Compaction | Background or blocking; background is default |
| Compaction cost note | Background may raise token use; no figure given |
| Auto Review | Evaluates commands for inappropriate behavior before execution; off by default |
| Time awareness | Experimental; SDK notes do not specify calendar vs elapsed-time scope |
| Session review | Programmatic inspection of agent sessions after completion |
| Tool settings | Tools can be updated mid-session |
| Key management | Organization external key management; callable keys scoped per request |
| API status | Public beta since September 10, 2026; beta client path and header |
Compaction is the harness summarizing earlier conversation so a long agent session fits the context window and keeps going, instead of stopping when the window fills.
Adopt SDK 3.42.0 without breaking a live agent
1. Upgrade in staging first — A drop-in upgrade for most projects, but beta behavior still deserves a staging pass.
2. Test both compaction modes — Run your longest real task under both modes; compare token use and result quality.
3. Opt into Auto Review deliberately — It starts off. Turn it on for command-executing agents, and measure false positives before promising safety.
4. Add session review to your audit trail — Inspect completed sessions programmatically instead of scrolling transcripts.
5. Keep acceptance outside the chat — Task, config and result live in your records, not in compactable context.
How to profit from it
Harness upgrades are consulting gold: every team running agents in production needs someone to adopt them safely.
- Sell an Agents API health check. Audit a client’s agent setup against SDK 3.42.0: compaction mode, Auto Review status, session review coverage, and key scoping.
- Offer a compaction A/B test. Run the client’s longest workflow under both compaction modes and report token cost versus quality.
- Build audit-ready agent services. Session review plus external acceptance records is a compliance story for regulated and enterprise teams.
- Do not oversell Auto Review. It is off by default and new. Present it as a safety layer to test, not a guarantee that dangerous commands stop.
- Bundle the Nov 30 migration. Teams moving off Agent Builder (see our Evals migration guide) need the Agents SDK anyway. One engagement covers both deadlines.
Sources
OpenAI Python SDK 3.42.0 release notes as covered by DEV Community and Emergent, citing the SDK repository. Check the SDK release notes themselves before changing a production agent’s settings.
Frequently Asked Questions
What is OpenAI Python SDK 3.42.0?
The October 31, 2026 release that adds agent context compaction settings, session review, experimental time awareness and organization key management.
Is Auto Review on by default?
No. Auto Review starts off and must be opted into. Background compaction, by contrast, is on by default.
What does background compaction do?
It summarizes long agent sessions in the background instead of blocking the agent, at the possible cost of higher token usage.
What is Auto Review?
A check, similar to Codex, that evaluates the potential for inappropriate behavior before an agent executes commands.
Is the Agents API generally available?
No. It remains in public beta, launched September 10, 2026, via a beta client path.
Keep Reading
Want more tools that actually pay for themselves? Browse the AI Tools section for the latest launches, pricing breakdowns, and profit playbooks — and check the official source linked above before you buy or deploy anything.