System prompts for production agents: the format that survived 3 redesigns (opens in new tab)

System prompts for production agents: the format that survived 3 redesigns The right way to make an agent do less is to write more in the system prompt. I know that sounds backwards. Let me explain. The first instinct when an agent goes off the rails is to cut its permissions at the application logic level. Validations, guardrails, filters over the response. But the problem usually lives earlier: the model doesn't have a clear contract for what's expected of it. And without a contract, it opt...