StudioTested LangGraph, Pydantic AI and OpenAI Agents SDK: a crash after approval ran the approved action twice
Original title: Approval, crash, resume: the approved action ran twice in 3 agent frameworks. How do you handle this?
In 30 runs per framework with fake models and real process kills, an approval followed by a crash executed the approved payment twice every time in LangGraph, Pydantic AI and the OpenAI Agents SDK. The author says this is at-least-once execution rather than a framework bug (all three projects were notified before publication), and suggests an external record, an idempotency key stable across resumes, and checking approvals against exact arguments.
The author ran controlled experiments with fake models and real process kills, and offers a checklist including idempotency keys that transfers directly to production agents.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.