Lately I've been messing around with AI agents to let them run on the chain by themselves, feeling like releasing a little robot outside the spaceship to clean the windows... It's quite romantic, but honestly, many steps still require human oversight.

For example, with permissions, I only give the minimum amount for signing/authorization. No matter how smart the agent is, it can still be led astray by prompts or accidentally click on strange contracts; then there's routing and slippage, where on-chain surprises happen all the time—MEV, sudden liquidity withdrawals—so it executes as planned, and when you look back, it's just a face full of question marks. And risk control—when to stop, when not to chase—agents can easily mistake "execution capability" for "faith."

Recently, I’ve also been complaining that on-chain data tools and tagging systems are a bit lagging, even misleading at times, so I’m even less willing to trust "tags = truth" to them. At most, they serve as a radar; the final decision still rests with humans. The same goes for trading psychology—I’m now more like practicing not to be led by noise: letting the agent run the process, while I just stick to rules and boundaries... Anyway, that’s how it is for now, slowly refining.