AI agents double-charged in 70 of 600 trials when payment tools time out
Soggy-Somewhere-8985 · reddit · 2026-10-29
The author ran the six provider-published payment toolkits (PayPal, Square, Adyen, Razorpay, Paddle, Coinbase AgentKit) unmodified against local fakes that commit the money, then respond late or drop the connection. Six models (two each from OpenAI, Anthropic, DeepSeek) then decided what to do across 600 trials — no real money involved.
Key findings:
- No toolkit sends its own idempotency key on money calls, so a retry registers as a brand-new refund or charge.
- Money moved twice in 70 of 600 trials.
- A "check before retrying" system prompt halved the rate (47 → 23 of 300).
- One model caused 50 of the 70 failures; two others went 0-for-100.
- No answer ever signals an unknown outcome: Adyen returns "Failed to refund payment. Error: Unknown error" for refunds that actually succeeded.
Code, pinned versions, fakes, and per-trial records are public on GitHub (terminalstate/double-take).
More from coding & agent
- Microsoft launches Copilot harness in Foundry Agent Service, built by GitHub team — lee_stott · 2026-10-29
- BLA hardens apps against LLM reverse-engineering agents, tested against REA — Promptmethus · 2026-10-29
- If Vibe Coding Builds Software On Demand, What Is a Fintech Vendor For? — mikeflache · 2026-10-29
- Spec first, code second: one workshop experiment shows agents build the wrong app — Al_Grigor · 2026-10-29
- Veteran engineer: juggling 17 projects with agents, deep thinking is a muscle I'm losing — rudrank · 2026-10-29
- Free MCP server makes ChatGPT/Claude cite datasheet page numbers for every component value — Winter-Drop-1334 · 2026-10-29