According to Andrew Ng (@AndrewYNg), the single most important factor influencing the speed of progress in building AI agents is a team's ability to implement disciplined processes for evaluations (evals) and error analysis. Ng emphasizes that while it might be tempting to quickly address surface-level mistakes, a structured approach to measuring system performance and identifying root causes of errors leads to significantly faster, more sustainable progress in developing agentic AI systems. He notes that traditional supervised learning offers standard metrics like accuracy and F1, but generative and agentic AI systems pose new challenges due to a much wider range of possible errors. The recommended best practice is to prototype quickly, manually inspect outputs, and iteratively refine both datasets and evaluation metrics—including using LLMs as judges where appropriate. This approach enables teams to precisely measure improvements and better target development efforts, which is crucial for enterprise AI adoption and scaling. These insights are shared in depth in Module 4 of the Agentic AI course on deeplearning.ai (source: Andrew Ng, deeplearning.ai/the-batch/issue-323/).