OpenAI Agents API Adds Auto Review, Async Compaction, and Cross-Session Prompt Caching

Auto Review puts a check in front of command execution

When developers wire an agent up to a terminal or external tools, the riskiest moment is the handoff between reasoning and actually running a command. In its October 30 weekly shipping rollup, the OpenAI Developers team introduced Auto Review for the Agents API: an automated review step that evaluates commands before execution takes place.

Instead of requiring every application team to build a custom interceptor loop from scratch, Auto Review sits directly in the command execution path. Pairing that pre-execution check with the rollup's tool updates tightens how agents invoke actions during multi-step runs without forcing a human to approve every benign read operation.

Async Compaction and cross-session prompt caching tackle latency and token bills

Long-horizon agents accumulate tool outputs, file reads, and intermediate traces until the context window fills up and slows down. Previously, compacting that history often meant pausing the agent's main loop while a summarization pass ran. Async Compaction moves context compression into the background as an asynchronous job, keeping the agent responsive while trimming older state.

On the cost side, the update enables automatic system prompt caching across sessions. Production agents typically carry large, static system instructions and tool definitions that repeat across thousands of separate user sessions; caching that system prompt across sessions trims redundant input token billing and speeds up the first turn of each new session.

• Auto Review

  Screens commands automatically prior to execution so unsafe or unintended actions are caught before they run.

• Async Compaction

  Compresses growing session context in the background without blocking the active agent loop.

• Cross-session prompt-cache

  Automatically caches system prompts across sessions to lower token spend on repeated agent startups.

• Time awareness & tool updates

  Adds experimental temporal awareness alongside refinements to built-in Agents API tools.

Experimental time awareness for agents that don't live in a vacuum

Standard language model calls treat every turn as a timeless snapshot unless the developer manually injects timestamps into every message. The October 30 update introduces experimental time awareness inside the Agents API, giving agents a sense of temporal context as tasks unfold.

Taken together, the five items in this weekly release focus less on flashy demos and more on the plumbing required to keep autonomous loops running in production: safer command execution via Auto Review, non-blocking memory management via Async Compaction, temporal grounding, updated tools, and lower per-session overhead through automatic system prompt caching.