Lately, I’ve been tinkering with letting an AI agent do on-chain work by itself, and the more I do it, the more I feel that “full automation” is a bit mysterious… To put it plainly, it can run the workflow, but some pitfalls still need someone to stand guard. For example, when it comes to signing, no matter how smart the agent is, it’s hard for it to judge whether this is being induced to grant authorization—especially with cases involving infinite allowances, contract upgrades, and a bunch of hops like cross-chain bridges. Even if the experience is smooth, someone still has to take a final look. There’s also the anxiety of repeated calendar-driven pressure around things like unlocks or staking unlocks: an agent may sell decisively according to the rules, but it’s very human to decide whether the rules themselves should be temporarily adjusted (sell less, split into batches, or just don’t move at all). My current approach is more like this: the agent compresses the steps into a single workflow plus a pre-run outcome, and I either give the nod or hit pause. For now, let’s do it this way.