I tried once to have an AI agent run processes on the blockchain: first scan the contract, then place an order, and finally record the transaction in a table. Honestly, everything that can be automated goes smoothly, but when it comes to the step of "whether to sign," I still need a human to oversee—especially for authorization limits, target addresses, and contracts where you can't tell at a glance if fees will be deducted or callbacks will occur. The agent seems very smart, but when parameters are a bit tricky or routing jumps twice, it easily risks "taking on the risk" just to complete the task.

There's also a part that must be done manually: reconciliation of results. It can tell you if a transaction succeeded, but it's hard to automatically verify where the funds came from, where they went, or if there's an inexplicable transfer out. Recently, I saw someone stumble again on the inflation + studio + token price spiral in blockchain games. Actually, many times it's not that you can't play, but that the accounts aren't properly reconciled, or authorizations haven't been revoked, which eventually turns into "the more you run, the more you lose"... Anyway, my current principle is: the agent handles the chores, and the person is responsible for signing and review. Taking it slow is more reliable.