A Field Guide to AI Documentation: Model Cards, Eval Reports, Agent Cards, and More

As AI systems move from deterministic code to probabilistic models and autonomous agents, traditional documentation like READMEs is no longer sufficient. James Anderson’s guide explores the emerging landscape of AI documentation, which functions as a set of 'trust artifacts' rather than just technical manuals. The article categorizes these new standards into four groups: model and data documentation (Model Cards, Datasheets, System Cards), evaluation reporting (Eval Reports, Benchmark Cards), agent-specific documentation (Agent Cards, Policy Cards, AGENTS.md), and governance compliance (Audit Trails, Transparency Reports). Because AI behavior can vary, these documents focus on ingredients, measurement, boundaries, and history to ensure transparency and accountability. As regulations like the EU AI Act and emerging frameworks mandate these practices, developers must treat documentation as a critical skill for building trustworthy systems, shifting the focus from simply asking 'does it work?' to proving how and why it functions.