Testing 30 payment-enabled agents: 5 money-loss bugs where failure looks like success
Tiny_Associate_8254 · reddit · 2026-10-26
A developer who spent weeks testing 30 payment-enabled agent projects (x402 clients, per-call LLM routers, wallet toolkits, budgeted autonomous agents—always on test tokens) reports that almost none of the failures looked like failures: the agent said "done", logs looked fine, and the ledger disagreed. The five recurring bugs: (1) lost replies after settlement cause double payments; (2) spend caps only count settled spend—three concurrent requests blew through a 0.012 cap at 0.024; (3) a wallet toolkit reported reverted transfers as "transferred"; (4) an int-vs-string status comparison marked every successful approval as failed after an unlimited allowance was granted; (5) an x402 client signed a fresh payment authorization on every retry, leaving multiple redeemable authorizations per request. Fixes: treat "unknown" as its own state and check the ledger, reserve budget before sending, read real receipts not mocks, one authorization per obligation, and put rules in code rather than prompts. All bugs are fixed or publicly disclosed.
More from coding & agent
- Fudan and Tencent unveil DraMaze, a multi-agent pipeline for diverse microdrama scripts — Siyang He · 2026-10-26
- Alibaba's MultiAgent-OPD trains hierarchical multi-agent systems end-to-end with on-policy distillation — TaoLiveAIGC · 2026-10-26
- Why Apple's folding iPhone shows black bars: apps built with older tools are deliberately boxed — aakashgupta · 2026-10-26
- License gotchas when picking an open source OCR model for commercial use — TangeloOk9486 · 2026-10-26
- Mistral ships Voxtral 3: speech models and Voice Agent SDK across 50+ languages — AxSaucedo · 2026-10-26
- skillEd: auditable agent evals via CLI and MCP, evidence over agent summaries — MiserableFlatworm337 · 2026-10-26