Work agents require a system-level perspective combining foundation models with runtime infrastructure, and their reliability depends on five core capability families and joint optimization of both the model and execution harness.
This paper presents a comprehensive framework for understanding work agents—AI systems that execute complex tasks across tools and applications.