AI agents are evolving into sophisticated systems known as agentic workflows, where multiple agents or a single complex agent with various capabilities coordinate to accomplish tasks with minimal human intervention. These workflows are dynamic, capable of adapting to changes in real-time, unlike traditional automation, which is more static and linear. The complexity of agentic workflows necessitates robust evaluation metrics such as task adherence, tool call accuracy, reasoning quality, and recoverability, as these systems are prone to errors that are harder to detect. Evaluating these workflows goes beyond ensuring task completion to verifying the correctness and efficiency of the entire process. Methods for evaluation include human-in-the-loop assessments, automated checks using AI models, and frameworks like AAEF that log and audit tool usage for compliance and improvement. Common pitfalls in evaluating these workflows include over-reliance on static benchmarks, ignoring process-level evaluations, and inadequate logging. Best practices for building agentic workflows emphasize modular design, real-time observability, and a mix of human and machine evaluations to ensure accuracy and adaptability in decision-making processes.