[Release] openai/openai-agents-python v0.21.0: v0.21.0

OpenAI Agents Python v0.21.0 is a minor release introducing provider-neutral testing APIs and OpenAI Python v3 compatibility. Key additions include new testing utilities for deterministic Agent, Sandbox, Realtime, and Voice workflows without provider requests, updated HTTPX2-aware request handling, and numerous hardening fixes across core, MCP, Sandbox, Realtime, and Voice modules. The release maintains backward compatibility while improving state isolation, error handling, and validation strictness.

Key Points

• •Added `agents.testing`, `agents.realtime.testing`, and `agents.voice.testing` utilities for deterministic, provider-neutral workflow testing without external API calls
• •Updated OpenAI provider compatibility to `openai>=3.0.0,<4` with full HTTPX2 support for request, response, transport, and exception handling
• •Hardened RunState interruption snapshots and isolated state checkpoints to prevent cross-request contamination
• •Improved MCP lifecycle management with snapshot isolation, lifecycle result isolation, and configurable retry backoff ceilings
• •Enhanced Voice validation with stricter checks for invalid channels, frame rates, incomplete multichannel frames, and non-finite audio rates
• •Fixed recursive agent-tool approvals and max-turn finalization session semantics for proper workflow resumption
• •Added Sandbox Runloop support for existing secrets and tightened view-image path grants for improved security
• •Improved streaming cleanup and sensitive error redaction to prevent data leakage in output-type validation errors
• •Fixed parallel tool call alignment with converted tools across core and extensions modules
• •Enhanced test infrastructure with sequential anchor application and improved test span ordering by start sequence

Workflow Diagram

Concepts