AI agents that move money need a safety check before action

An autonomous agent that trades Solana memecoins faced its biggest problem right before execution, not during decision-making. In early tests, about 42% of the tokens it bought collapsed to almost zero. Some tokens looked safe on the surface: the contract was verified, the liquidity pool was burned, and there was no obvious large holder waiting to sell.

The real danger sat outside normal language reasoning. Groups of wallets had been funded from the same source, bought in the same block, and were set up to sell into the next buyer. The fix was a required tool call before the agent signed a trade.

That tool returned a 0–100 risk score and flags for shared funding sources, same-block buying groups, and deployers with repeated scam history. If the score passed a set limit, the agent skipped the trade. With that check and an entry-timing rule, the collapse rate fell from about 42% to nearly zero.