Late at night, I came across a bunch of posts about AI Agents automatically onboarding to the chain and executing trades, as if they don’t need human oversight... I actually feel more nervous watching that. To be honest, what Agents can really do is just click buttons according to the rules you give them, but the most problematic steps on the chain still require human oversight: who exactly was given the initial authorization, is it an unlimited grant; has the contract address been swapped or cloaked; what permissions are being transferred in that string of “small unreadable text” in the signature popup. Especially when dealing with routing/aggregator protocols that jump from one to another, Agents might only chase the optimal path, but humans need to watch out for the worst-case scenarios: slippage, MEV, where assets go during failed rollbacks. Recently, many people are pushing narratives, but I’d rather see who’s willing to tighten default permissions, fully whitelist, and make revoke options more visible... Anyway, my usual approach is: Agents can run, but I only give them “a range where mistakes can happen” in terms of permissions and limits.