OpenAI Agents API ships Auto Review, async compaction, experimental time awareness and mid-session tool updates

Background and the problem

OpenAI's Agents API entered public beta on September 10. It turns the Codex agent harness into a managed service: OpenAI handles session orchestration, context compaction and recovery, while developers bring their own tools, MCP servers and sandboxes. The usual pain points for hosted agents are agents running dangerous commands, stalls when long sessions compact their context, models not knowing how long they have been working, and tool lists that cannot change once a session starts. This week's updates target most of these.

Architecture and how it works

According to a post on X by Steve (@stevendcoffey, whose bio reads "API things @openai"), shared by the official OpenAI Developers account, five things shipped this week. First, Auto Review, which works like the Codex mechanism and evaluates commands for potential bad behavior before executing them. Second, async compaction, which lets context compaction happen in the background to speed up sessions. Third, experimental time awareness, which gives the model a sense of time passing so it can better act on instructions like "timebox CI fixes to 10 minutes." Fourth, a session's tools can now be updated at any time. Fifth, system prompts are now cached automatically across all sessions to save money and time.

Benchmarks and measured results

OpenAI has published no benchmarks, and gave no figures for how much faster async compaction is or how much automatic caching saves. As of writing, the latest Agents API entry in the OpenAI API changelog is still October 16 (Artifact SDK in hosted sandboxes and Ultrafast mode), and these five items are not listed yet, so check later docs for exact parameter names and defaults. For reference, earlier iterations included computer use on September 29 and sandbox prewarming with active-turn billing on October 9.

Practical guide for developers

Teams already running coding or ops agents on the Agents API should look at Auto Review first: it adds a check before commands run, which suits setups that grant more autonomy, though sandbox isolation and human approval are still worth keeping. For long sessions, watch how time to first token changes with async compaction. Multi-session apps with long system prompts can compare input costs on the usage page before and after cache hits. Time awareness is experimental, so keep a hard timeout on the application side for time-boxed tasks. See the Agents API docs, the compaction guide and the prompt caching guide for details.