These past few days I’ve seen a bunch of narratives about AI Agents plus on-chain interactions, hyping it as “we won’t need humans anymore”… I’m a pretty straightforward, cautious type of person: automate as much as possible, but some steps still need a human to stand guard—otherwise you’re just outsourcing the risk to luck.

For example, with authorization, no matter how smart the Agent is, it may still cut corners and give you an infinite approve. In plain terms, all it takes is one phishing attempt or one fake contract, and you’re in trouble. Another example is parameter boundaries: slippage, gas, transaction routing—once the on-chain environment changes, it may start “executing with confidence,” and then you get hit with MEV or end up with the trade filled at some wildly unreasonable position. And for cross-chain/bridges, if any step stalls, that’s not something AI can “figure out”; in the end, you still have to confirm whether you should stop, withdraw, or switch the route. Automated trading is the same: the strategy logic can run on its own, but in abnormal cases (sudden liquidity collapsing, the pool being altered, a contract being upgraded), someone has to spot immediately that “something is off.”

It feels like the people who truly obsess over security are doing very basic things: whitelists, minimum permissions, limits, and options to roll back or pause. Meanwhile, those who love the hype like to talk about “fully automated closed-loop systems.” What I’ve learned isn’t a trick—it’s this: the more automation looks like it will be hassle-free, the more you need to put “can I stop it at any time?” first.