Approved action ran twice in 3 agent frameworks: crash-resume test
KaleidoscopeNew9259 · reddit · 2026-11-06
The author stress-tested human-in-the-loop flows in LangGraph, Pydantic AI, and the OpenAI Agents SDK.
Setup
- 30 runs per framework, fake models + real process kills
- Scenario: crash right after the approved action executes, before the framework saves progress
Findings
- All three frameworks reproduced "one approval → two executions" (e.g., double payments)
- Root cause is at-least-once execution semantics, not a framework bug; all three projects were notified before publication
- Keeping a record outside the framework reduced it to a single run
Checklist for your own agents
- Does your payment/email API get an idempotency key consistent across resumes?
- After a mid-action crash, does the agent retry or stop and ask?
- Is the approval checked against the exact arguments about to run?
The author maintains a small open-source library addressing this and will share the write-up, scripts, and data in the comments.
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