Dev.to
7/3/2026

The original headline is: "OpenAI Agents SDK 0.13 to 0.17: Three Breaking Changes You Will Hit"
Original: OpenAI Agents SDK 0.13 to 0.17: Three Breaking Changes You Will Hit
Short summary
OpenAI's Agents SDK (0.13→0.17) introduced three breaking changes: model refusals now raise exceptions, the default model switched from GPT-4.1 to GPT-5.4-mini, and Sandbox Agents enforce path security. All three can silently break production agents without error at upgrade time. The guide provides a migration checklist and safe upgrade pattern with explicit model pinning and refusal handlers.
- •Model refusals raise exceptions; add error handlers to avoid crashes
- •Default model switched from GPT-4.1 to GPT-5.4-mini with different quality/token tradeoffs
- •Sandbox Agents now enforce path security restrictions; use path grants or restructure files
Generated with AI, which can make mistakes.
Is this a good recommendation for you?



