OpenAI's Agents API now reviews risky commands before they run, and can be told to timebox a task to 10 minutes
OpenAI's Agents API team shipped a batch of updates this week. Auto Review checks the commands an agent wants to run for bad behavior before executing them, the way Codex already does. Compaction can now happen in the background so long sessions don't stall, and an experimental time-awareness mode gives the model a sense of time passing, so instructions like "timebox CI fixes to 10 minutes" actually work. You can also change a session's tools at any point, and system prompts are now cached across sessions automatically.
Why it matters
These are the pieces long-running agents keep tripping on: unsafe commands, sessions that freeze to compact, and no sense of a deadline. Automatic prompt caching also trims the bill.
Who it's for
Developers building agents on the OpenAI Agents API, especially CI, ops and coding agents that run for a long time.
How to try it
- Update your SDK (for Python, openai 3.42.0 or later).
- Turn on session review so commands are checked before they execute, and enable async compaction for long sessions.
- Try the experimental time awareness with an instruction like "spend at most 10 minutes on CI fixes, then report back".
- Keep your system prompt stable across sessions so the automatic cache can save you tokens.
Plan & availability: Agents API (public beta) on the OpenAI platform, billed at normal API rates. The openai-python 3.42.0 release (October 31) adds the compaction, review, tool and time-awareness settings. Time awareness is experimental. Announced by OpenAI's Agents API team and the OpenAI Developers account on October 30, 2026.
Source: Steve Coffey (@stevendcoffey), OpenAI on X ↗
Written by AgentStackList from the original post. Checked Oct 31, 2026. Features, plans and availability can change, so confirm on the source before you rely on it.
openaiagents-apiauto-reviewcompactiondevelopers
More ChatGPT tricks
- ChatGPT finally lets you stay signed in to a work and a personal account and switch in one tap · Oct 31
- OpenAI Private Intelligence: zero data retention, private inference and approval rules for business agents · Oct 30
- An OpenAI model slipped onto the live internet during training. Monitors caught it in 15 minutes · Oct 29
- OpenAI backed down: the Pro 200 cut is cancelled and Pro 100 and 500 get double the Codex usage · Oct 29
Agent Stack Weekly · free
Get the week's best AI tricks
Limit and pricing changes, new MCP servers, and setup fixes for Claude Code, Codex, Cursor, Copilot and Antigravity.
No spam. Unsubscribe anytime. Read past issues