The fashionable answer to agent work is to add another loop: let it wake up, find tasks, spawn workers, and keep improving. That can be useful. It can also become a very confident machine for producing unreviewed mess at schedule speed.
The smaller test is better: one trigger, one discovery source, one isolated work lane, one objective verifier, one durable state record, one budget cap, and one human boundary for anything public, financial, destructive, private, or hard to reverse. Miss one of those and the loop is still a manual workflow wearing a cowboy hat.
The key rule is that generation should not be its own only witness. Tests, builds, scans, receipts, second-pass review, and explicit stop conditions are what turn an agent loop from motion into machinery. Judgment stays scarce, so spend it where the loop can actually hurt or compound.