Crash test: LangGraph, Pydantic AI and OpenAI Agents SDK all double-charge payments
KaleidoscopeNew9259 · reddit · 2026-11-06
Developer Salah Makboul ran a crash test across LangGraph, Pydantic AI, and OpenAI Agents SDK: kill the process right after an approved tool payment, then resume per each framework's docs. 30 runs per framework (Oct 21–Nov 5, latest versions daily).
- Without a guard: all three frameworks paid twice every run
- With an external record: exactly one payment
- Tampering the saved call to pay 9999 to "mallory": LangGraph and Pydantic AI paid it; OpenAI's SDK refused on its own
The finding isn't a bug but inherent at-least-once execution: doing the work and saving that you did it are two steps. All three projects were notified 14+ days ahead—OpenAI added docs, Pydantic AI reproduced it, LangGraph docs already warn.
Mitigations: provider-enforced idempotency keys, treating "maybe happened" as unknown instead of retrying, and validating approvals against exact tool arguments. Write-up, scripts, and per-run data on GitHub (author maintains the free Call4Agents library).
Related event: Agent Frameworks Double Payments After Crash, Tests Show(2 posts)→
More from coding & agent
- Drew Breunig's QCon Talk Breaks Down Prompt Debt and Four Context Failures With Concrete Fixes — dbreunig · 2026-11-06
- Dev essay: why a runaway intelligence explosion isn't the risk — 'early onset decoherence' is — round · 2026-11-06
- Sardine launches Know Your Human protocol to verify humans behind shopping agents — astrange · 2026-11-06
- Indie dev cuts 5,672 lines rebuilding trip planner's Days screen into one-drag editing — SeanOliver · 2026-11-06
- LangChain details optimizing browser use for deep agents; design agent fixes its own blog formatting — LangChain · 2026-11-06
- Baker Lab + NVIDIA unveil RFD4-Proteina: agents run protein design tool calls on GFP — AllThingsApx · 2026-11-06