Writing Evals for Agentic Systems as a 4 Step Loop (sgaud.com)

A recent article outlines a comprehensive four-step evaluation process designed specifically for agentic systems, such as large language models (LLMs). The significance of this approach lies in its recognition of the non-deterministic nature of these systems, where outputs can vary widely. The process begins with storing a versioned snapshot of the agent's operations, followed by a manual evaluation from a subject matter expert (SME). This chain of evaluations focuses on capturing both intermediate behaviors and outcomes, establishing a framework for feedback and identifying recurring errors. Subsequently, a dataset of SME evaluations is created, leading to the formulation of a scoring rubric based on observed failures. The final step introduces an LLM-as-a-Judge that leverages this dataset to assess future agent runs, either during execution or in post-analysis. This methodology not only streamlines evaluations by reducing SME workload but also necessitates regular updates to the judging model to prevent evaluator drift. Overall, this structured approach enhances the reliability and scalability of evaluations, marking an important advancement in the assessment of AI systems that operate in uncertain environments.