Agent Engineering: A New Discipline

Agent engineering is proposed as a new discipline for refining non-deterministic LLM systems into reliable production experiences. It is an iterative, cyclical process of building, testing, shipping, observing, and refining, where shipping is a means to learn rather than a final goal. This discipline combines skills from product thinking, engineering, and data science to manage the unpredictability of AI agents. The need for this approach arises because while agents can now handle complex workflows, their unpredictable behavior creates unique challenges that traditional software development practices cannot solve.