OpenAI Agents SDK Python v0.22.0
OpenAI Agents SDK Python v0.22.0 introduces runtime hardening, redacting guardrail-blocked tool outputs from replay state, raising ModelBehaviorError on failed/incomplete non-streaming responses, and strictly rejecting conflicting provider client options.
Why this signal matters
Version 0.22.0 focuses on runtime hardening and provider configuration enforcement. Key updates include: redacting terminal function-tool outputs rejected by guardrails from replay and persisted SDK state; raising ModelBehaviorError for non-streaming responses with terminal status 'failed' or 'incomplete'; rejecting conflicting options when OpenAIProvider receives an explicit client (requiring organization/project to be passed directly to AsyncOpenAI ); isolating token usage accounting between independent RunState checkpoints while retaining nested-agent aggregation; expanding handoff(agent) targets in generated visualization graphs; and clarifying shallow-copy behavior in Agent.clone() and RealtimeAgent.clone() documentation.
Actionable summary
OpenAI Agents SDK Python v0.22.0 enforces stricter OpenAIProvider validation by rejecting conflicting options when an explicit client is provided. Additionally, it redacts blocked tool outputs from replay state, raises ModelBehaviorError on terminal failed/incomplete responses, and isolates usage across RunState checkpoints.
- Agent usefulness
- 82/100
- Confidence
- 96%
- Canonical data
- JSON + Markdown
What builders should check
- If combining an explicit OpenAI client with organization or project configurations, migrate those settings directly into the AsyncOpenAI client instance.
- Update error handling to catch ModelBehaviorError for non-streaming responses ending in terminal failed or incomplete states.
- Review state replay and persistence logic to account for the redaction of tool outputs rejected by output guardrails.