humanlayer / 12-factor-agents

What are the principles we can use to build LLM-powered software that is actually good enough to put in the hands of production customers?