[Release] openai/openai-agents-python v0.20.0: v0.20.0

OpenAI Agents Python v0.20.0 is a minor release introducing GPT-5.6-Luna as the default model, adding support for MCP Python SDK v1 and v2, and enhancing RunState with durable pending input staging. The release includes significant improvements to sandbox mount validation, realtime transcription settings, and numerous bug fixes across tool guardrails, serialization, and session management. Applications with custom MCP HTTP transports must migrate to httpx2 or pin mcp<2 to avoid breaking changes.

Key Points

• •Default model changed to gpt-5.6-luna; explicit models and OPENAI_DEFAULT_MODEL env var still take precedence
• •MCP Python SDK v1 and v2 both supported across stdio, SSE, and Streamable HTTP transports with potential breaking changes for custom HTTP auth
• •RunState.add_input() enables staging durable user input before resumed model calls with guardrail and persistence support
• •Sandbox mount validation now requires explicit credential-exposure acknowledgements while preserving redacted error contracts
• •Realtime input transcription supports GA settings for gpt-live-transcribe, gpt-transcribe, and gpt-realtime-whisper models
• •Fixed queue consumer deadlocks and improved streamed guardrail result publishing on failure paths
• •Enhanced tool approval binding to concrete invocations and normalized optional fields in tool invocation identity
• •Improved serialization handling for containers of models, dataclasses, and non-text MCP content blocks
• •Added schema validation to reject unsafe recursion depth and unsafe strict-schema ref siblings
• •Fixed RunState resume behavior to preserve shell outputs, guardrail session state, and tool approval context

Workflow Diagram

Concepts