Recently, I’ve been watching a bunch of AI agents “automatically doing work” on-chain—it’s pretty cool, but let’s be honest: someone still has to cover a few key steps. First is authorization/limits. Once an agent gets excited, it may grant you unlimited approvals, and if anything goes wrong with the contract later, people are going to be in trouble. Second is routing and slippage. The MEV crowd doesn’t care about your feelings—once a transaction is exposed, it could get snatched, and the agent only looks at the quote, not the battlefield. Third is cross-protocol yield stacking—especially setups involving re-staking and shared security. The “Russian-doll” (nested) criticism isn’t without reason either; the risks don’t add up in a simple linear way, and the biggest blowups often hide in some underlying layer you thought was stable. My approach is still: manually confirm the key signatures, periodically clear permissions/allowance windows, and use some protective tools (at least don’t go in naked). The agent handles the bricklaying—I handle the brakes… I’m going to get to work.