These days, I've been seeing everyone talk about AI Agents automatically running strategies on the chain. Honestly, to make it truly "fully automated," humans are still the ultimate fallback. The most prone-to-issues part of on-chain interactions: authorization (approve with large amounts at once), cross-chain/bridges, and contract upgrades/proxy contracts—these are the parts that look unchanged but actually have different logic. Agents might act based on "saving gas/fastest execution," but when encountering block delays, sudden slippage widening, or route congestion (like being squeezed), they might not know when to stop.

Additionally, when a major public chain undergoes an upgrade/hard fork/maintenance, I don't know if the ecosystem will migrate, but thinner liquidity and erratic quotes are common. During such periods, automation is more like an accelerator, not a steering wheel. There are many tutorials, but I prefer to see "failure case reviews" and "permission management" discussions—don't just go all-in and hand everything over to the robot right away. For now, staying alive is the most important.