Recently, I've been looking at AI agents interacting on-chain, which is indeed more convenient, but my obsessive-compulsive tendencies still require leaving a few "fallback" options for humans. The signing step must be confirmed by a person: authorization scope, limit, whether an unlimited allowance was given, if something looks off, revoke it. Also, with slippage/ routing, the agent is good at "helping you execute trades," but when liquidity is thinner or there are more traps, you only realize after the trade is done that you've been taken quite hard... I’d rather double-check a couple of times.

And then there's the strategy-level switches: when to stop, when to reduce positions. The agent will only follow the rules; if the rules are written incorrectly, it keeps making mistakes. In blockchain games with inflation and studio-driven price spirals, the agent can also "diligently execute," but the more it works in the wrong direction, the worse it gets. Anyway, my current habit is: let it handle execution and monitor the market, but I still review the key permissions and risk control thresholds myself to keep my confidence steady.