Source · release

OpenAI Agents SDK v0.24.0-alpha.8

Summary

The eighth preview of OpenAI Agents SDK 0.24 (2026-11-16) adds human approval of Python tool calls, support for model clients other than OpenAI's, realtime voice in sessions, and checkpoints that keep budget usage. It is a preview installed from the release's wheels, and it has several upgrade steps.

Key points

  • Tool approvals: review a Python tool's arguments and approve or reject the call before it runs (FunctionToolApprovalRequest, FunctionToolApprovalDecision); a decision can be remembered for the rest of the session.
  • Bring your own model client: the native tool loop works with your own async Responses or Chat Completions client, LiteLLM, AnyLLM, or a custom adapter.
  • Realtime voice: add voice to a session over a server WebSocket, WebRTC in the browser, or an existing SIP call (agents.realtime.native; the old API stays under agents.realtime).
  • Resumable work: pending runs and approvals can be checkpointed and restored with request or token budgets that keep what was already spent.
  • Tools and storage: ToolSearch defers eligible tool schemas, MCP tools can be picked by alias, SQLAlchemySQLiteSessionStore and CustomStorage (Redis, MongoDB, SQLAlchemy, Dapr) store sessions; read-only workspaces and native skills in execution environments.
  • Upgrade notes: back up saved state (alpha.7 cannot read alpha.8 checkpoints); replace bare "*" network rules with exact hosts or NetworkPolicy.allow_all(); ChatGPT Apps tools now need an explicit ChatGPTAppsMCP declaration. Local execution on Linux needs bubblewrap (bwrap).

Why it matters

Per-call approval and non-OpenAI model clients were the main gaps compared with other agent frameworks. Still a preview: wait for the stable 0.24 release for production.