Recently, I've seen everyone discussing AI Agents on the blockchain, and it does seem to save quite a bit of clicking and processes, but when it comes to actually spending money, it still requires human oversight. For example, during the authorization step, the agent might just want to "successfully complete the task," and might give you an unlimited credit limit. Frankly, it's hard for it to consider your risk concerns; or for cross-chain and token swapping routes, with so many options, it's easy to step into slippage, traps, or mysterious bridges. On-chain everything looks "successful," but the user experience ends up being a mess.

And with the current testnet incentives and points system, agents are best at following the process, but guesses like "Will the mainnet issue tokens?" ultimately depend on human judgment—whether it's worth it, whether to leave a trail, or whether to involve the main wallet. Anyway, my current approach is: run the agent with a small wallet, give it very limited permissions, sign critical transactions myself, take it slow if needed, at least I can sleep peacefully at night.