CAFE(S): Your Agent Is Only as Good as Its Context

AI coding agents have reshaped how software is built and delivered. But engineering leaders are quickly discovering that even the most capable models underperform when given low-quality context. This drives up token costs, opens organizations up to liability, and creates new sources of developer toil.

While context engineering focuses on assembling information for an agent, teams lack a shared standard to evaluate whether that context is actually fit for the task at hand.

This whitepaper introduces CAFE(S), a research-backed framework co-authored by researchers from DX, Capital One, GitHub, UVic, and Google. CAFE(S) establishes five durable properties for evaluating context quality: Clarity, Actionability, Fidelity, Efficiency, and Security. Taken together, these give engineering organizations a practical standard to diagnose context breakdowns, eliminate developer rework, and build the conditions for effective human-agent collaboration.