My OpenClaw agent runs while I'm away: what should stop it before a consequential action at 3am?
An OpenClaw user reflects on the risk of an always-on agent: it can run workflows, call APIs and move money, and the actions it takes at 3am are the ones nobody watches. The three cases are a reused approval applied to a different amount or payee, a payee off by a character or two, and a first-time payee with no limit set. The proposed fix is to bind each approval to the exact payload, make it single-use, and check it right before signing, with a deterministic gate and a human pause above a small threshold.