Treat LLM Agents Like New Bank Tellers: Five Controls That Transfer

A developer with a background in financial systems argues that LLM agents with tool access resemble new bank employees who are fast but cannot reliably distinguish genuine customers from fraudsters. The author suggests that long-established banking controls can be applied to these agents. Key recommendations include enforcing policy in code rather than in prompts, using a gate that checks each proposed tool call before execution, and treating retrieved document content as untrusted data to limit prompt injection. The piece also recommends detailed logging of every proposed tool call, including rejections, and using velocity checks to detect fraud patterns.