Lately, watching AI Agents run on-chain interactions has been pretty exciting; I feel they can do “repetitive grind-work” more steadily than people: monitoring funding rates, rolling over positions, switching pools on a schedule... but when it comes to the critical parts, you still need a human to backstop it. For example, if permissions are set too broadly, the contract address suddenly changes, or a route adds an unfamiliar pool one hop away—those “looks wrong at a glance” issues—an Agent might just keep clicking through the process and confirming; to put it plainly, it isn’t afraid of getting hurt.

And I get the complaints about how every new L1/L2 drops incentives and TVL just pulls in immediately—“mine, then sell” criticisms aside—I feel Agents are better suited for executing a retreat plan: pulling out on time, don’t drag it out into a long fight; but whether to enter or chase that rush of hype still comes down to the person handling their own emotions (I also want to go in—I’ve literally got my hand on the mouse). In any case, what I’m doing now is: let it run with small positions, but for large positions I must manually recheck—I'd rather move slower than let automation both create the problem and automatically take the blame.