I recently looked around at AI Agents performing on-chain interactions, and yes—they’re quite capable. They can automatically make markets, do bridging/arbitrage, and write good strategies that don’t need attention for days. But honestly, based on what I’ve observed, there are still a few steps that can’t be done without human oversight.

First, when a contract throws an exception or the chain gets congested, the Agent may end up in a dead loop or burning gas for nothing—someone needs to manually stop it, then restart. Second, when the underlying logic changes—such as a protocol updating its parameters or introducing a new governance proposal—the Agent can’t reliably decide on its own whether it should “reroute.” Third, once funds are spread out, if something goes wrong with a specific on-chain data (DA) layer and the data can’t be retrieved, the Agent will keep waiting blindly—people need to pull out quickly.

Modularity has been quite hot lately; developers talk about it enthusiastically. But as a user, to be honest, when I see the middle layers stacked on top of each other, it makes me feel a bit uneasy. If things get too complex and one module fails, the Agent won’t even know who to contact. For now, I still keep a fallback: large amounts go through manual handling, and smaller positions let the Agent run—earn slower but sleep well.