Someone asked me: since AI Agents can now go on-chain by themselves, what are humans still for as the safety net? My own feeling is this—when it comes to the moment of signing, you still need someone watching closely, especially for authorization, quotas, and the target contract. Once an Agent “gets it right on the side,” and sets you up with an unlimited approval, it’ll be too late to regret it later.

Then there are routing and slippage—yes, it may follow the optimal route, but if it runs into a pool being squeezed or liquidity suddenly thinning out, someone still has to make the call on whether to back out.

And for traps like contract upgrades and replacing an implementation via proxy contract changes—an Agent might not be suspicious enough to thoroughly check those.

Haven’t people also been complaining recently that on-chain data tools and the tagging system are a bit laggy and can even be misleading? In short: if you feed the Agent’s judgment into these tags, it can also be led off course.

Anyway, what I do right now is: let it handle the process, but for key actions, keep a “confirm button”—and I’ll treat myself as the last line of defense.