It’s raining today and traffic is terrible—the coffee in the car will cool off in a moment… It’s just the right moment to think about how everyone’s been talking up AI Agents for automatic on-chain actions lately. I’m definitely tempted, but I can’t help asking: where exactly do humans still have to be the safety net?

To put it plainly, what can mostly be automated on-chain is “execution,” but “intent” and “boundaries” still have to be handled by people. For example, when it comes to authorization/limits, no matter how smart the agent is, it still won’t take responsibility for the consequences if you accidentally authorize too much—like clicking/tapping the infinite-approval option by mistake. And for things like cross-chain operations and switching pools, the small traps—slippage, routing, and taxes/fees—ultimately end up being blamed on the person who signs with their wallet. Also, when there are contract upgrades, permission changes, or sudden pause switches, someone needs to verify the information source; otherwise the agent will run under old rules, and it’s only normal that it could go wrong.

Recently, in some areas, taxes and compliance measures keep tightening and loosening in waves, and expectations for deposits and withdrawals fluctuate accordingly—making it easier for people to think, “Let the machine do it for me, I’m in a hurry.” But the more such times come, the more you should slow down by half a beat: it’s fine to let the agent help you monitor the market and screen conditions, but when it comes to changing wallet permissions, moving large amounts, or interacting with unfamiliar contracts, I still prefer to look it over myself and then sign. I’d rather make a little less than wake up one day to find I’ve handed it a master key. For now, that’s it.