Before giving an AI a wallet, first check whether it can avoid duplicate payments. The challenge in agent payments is shifting from “able to pay” to “knows how to spend.” ① BlockRun released the Agent Money Benchmark, testing models’ performance in research tasks and adversarial transaction scenarios using real USDC payments on Solana. ② According to the project’s published results, seven of the ten tested models paid at least one fraudulent seller at least once, and six made duplicate purchases of the same data. These results come from the project’s own tests; generalizability still requires further replication. ③ The so-called “wasted spending” also includes the data and computational costs incurred for low-quality responses—not all of it should be counted as outright fraud. The test tasks, seller setups, and evaluation methods all influence the final numbers. ④ My focus is on product design: budget limits, duplicate payment checks, quote expiration dates, seller reputation, and result validation—all must form constraints before and after payment. On-chain settlement accuracy does not guarantee rational purchasing decisions. My own view: Agent commercialization is promising, but I care more about whether it can explain exactly what each payment bought. Reliably completing small tasks is more convincing than demonstrating a single automated transfer. For any tool requiring authorized funds, understanding permissions and budget boundaries should come before discussing scalability. NFA / DYOR @Solana_zh @solana @BlockRunAI