Measuring Agents in Production

A review of the December 2025 paper 'Measuring Agents in Production' (MAP), which surveyed 306 practitioners and 20 in-depth case studies across 26 domains. Key findings: 80% of production agent systems use predefined structured workflows rather than autonomous planning, 68% require human intervention within 10 steps, 70% rely on prompting off-the-shelf models rather than fine-tuning, 85% of teams build custom infrastructure instead of using frameworks like LangChain, and 75% skip formal benchmarking. The author draws a parallel to blockchain hype circa 2018, but concedes AI agents differ in that they do deliver real value—just in a far more basic, constrained, and human-supervised form than the hype suggests.