OpenAI's Agents SDK v0.24.0-alpha.6 preview adds session control and tool policies

The preview adds a way to submit work through a captured Turn, with steering, cancellation and queued prompts via Session.submit(). It also adds LocalStorage for reopening conversations from a local profile, plus NativeOutput and PromptedOutput for structured output. An optional max_retries setting performs bounded corrections without tools, and LocalStorage requires max_retries=0.

Tool handling gets several changes, including tool policies for inspecting arguments and results, per-submission tool selection, and native MCP argument replacement and result transforms. NetworkPolicy is added for native commands on supported Linux and macOS executors. Native OpenTelemetry export and agents.testing.ScriptedResponses for offline testing without a live model or API key are also included.

Several APIs are renamed or removed. Runner aliases are gone, Turn.wait() and Session.wait() are replaced by Turn.get_result(), and Session.interrupt() becomes Turn.cancel() or Session.cancel_all(). Subagent.submit() is renamed Subagent.start(), and ManualTool is folded into FunctionTool.

Agents built on this preview can persist and resume sessions, recover pending manual tool requests from checkpoints, and test flows offline with scripted responses. Code that calls the removed Runner aliases, Turn.wait(), Session.interrupt() or Subagent.submit() will need updating before adopting it. Because this is an alpha installed from GitHub wheels rather than PyPI, pinning the exact build is advisable and the API may keep changing.