These days I've been looking into AI agents interacting on-chain, and the more I look, the more I feel: automation is pretty appealing, but ultimately humans still need to be the fallback. A few days ago, I almost got caught out; the agent was supposed to switch to a different pool according to its strategy, but there was an old contract in the path, and the permissions/upgrade traces looked suspicious. I was lazy at the time and thought, "It's all verifiable on-chain anyway"... Luckily, right before confirming, I took a quick look at the authorization and contract history, and manually stopped it. Otherwise, I might have ended up with a situation where "the money wasn't lost, but the authorization was left open," which is pretty scary.

To put it simply, an agent can help me monitor large transactions, calculate slippage, and choose timing, but things like authorization scope, what the signature actually does, whether the contract has been recently modified, and whether to execute in extreme market conditions still need human judgment. Recently, people have been interpreting ETF capital flows and US stock risk appetite together to analyze market movements. I do that too, but I’m more afraid of being led astray by the narrative... Let the agent do its work, and I’ll stay a bit more suspicious.