These days, I've been looking at AI Agents running on the chain again, and I feel they are most like "people who are very quick with their hands but emotionally dull." Checking data, monitoring pools, placing orders according to rules are all quite steady, but when it comes to boundary cases, they need humans to back them up: for example, how much authority/signature to give, whether the contract address is a shell swap, where exactly the cross-chain step gets stuck, and whether to force through sudden slippage spikes. Honestly, Agents are good at execution but not very responsible for the "consequences."

Additionally, the community has recently been arguing about privacy coins, coin mixing, and compliance lines. Watching this, I become even more certain: when it involves identity, source of funds, or interactions with others, don't expect robots to make moral choices for you... They will only amplify your preset preferences, along with the blame. Anyway, my current habit is: let it prompt, filter, and simulate, but the key step is still to press it myself, to avoid self-educating after the fact by looking at transaction records.