Lately, AI Agents doing on-chain interactions have been pretty popular. I also tried letting it automatically perform some “tasks,” but I found that there are actually quite a few places where human oversight is still needed. The most critical step before signing: it can understand “authorization,” but it doesn’t necessarily understand “how much authorization to give, to whom, or whether it might conveniently package up your wallet keys and take them along”… In the end, I still have to watch the contract address and the allowance myself, and confirm once again before signing.

As for sorting and delay, the Agent is diligent, but on-chain rewards aren’t based on “who works harder.” If it gets stuck on a block, if slippage suddenly shifts, and if there’s an MEV slot ahead that cuts in line, then what was supposed to be “worry-free” turns into “post-event review.” During airdrop season, plus task platforms that are anti-sybil, it also feels like they’re giving Agents a set of hard problems. The points-based system makes it feel like clocking in for work. If you let it run automatically, the risk feels more like helping the platform test its risk controls; if you run it manually, it’s exhausting. In any case, my approach now is: automate when possible, but only let it do “check” and “simulate.” When it comes to the moment where actual money moves, I still do that part myself.