"Instruct autonomous agents to ask before destructive actions When running fully autonomous AI agents on live accounts, the speaker relies on a simple safety instruction rather than technical sandboxing: 'you can just tell it don't do any destructive actions — ask me first. You can do that.' While he acknowledges this relies on the model's judgment, Claude Opus in particular has proven reliable enough that this verbal constraint provides meaningful protection. Any knowledge worker delegating real account access to an agent should make 'ask before irreversible actions' a standing instruction in every prompt. Source: Swix, AI in Action, AI in Action Live Session"