Human Approval for AI Agents Using Step Functions Callback Pattern

The article argues that AI agents performing costly or irreversible actions need a human checkpoint before proceeding. It proposes using AWS Step Functions' wait-for-callback pattern, which pauses a workflow until a person responds, and can wait up to a year at no waiting cost. Actions are sorted by reversibility and cost, with small low-risk actions automated and larger ones routed to human approval. In the example, a bookshop's support agent may auto-apply credits up to 20 euros, requires team approval up to 500 euros with escalation after four hours, and is blocked entirely above 500 euros. The author stresses that these rules belong in the workflow rather than the prompt, since language models can be forgotten or manipulated, whereas choice rules cannot. The task token that resumes a paused step is treated as a sensitive credential and should be kept away from emails and the agent.