Researchers define five pillars for high-quality agentic context
A new framework from DX, Capital One, GitHub, University of Victoria, and Google identifies context quality as the primary bottleneck in AI agent performance. The authors argue that treating context as a core engineering discipline is essential to prevent compounding failures in autonomous workflows.
Researchers from DX, Capital One, GitHub, the University of Victoria, and Google have published the CAFES framework, a research paper that defines five properties for high-quality context in agentic workflows: clarity, actionability, fidelity, efficiency, and security. The paper posits that context quality is the central bottleneck in AI agent effectiveness, distinct from model capability or harness engineering. The authors identify specific failure modes caused by poor context, including 'weekend runaway,' where agents expand scope due to undefined completion criteria, and 'lost in the middle,' where agents overlook relevant data in large context windows. They also highlight prompt injection as a critical risk. The framework excludes 'accessibility' as a dimension, placing it as a prerequisite outside the five pillars. The authors propose that low-quality context creates a negative flywheel where bad outcomes feed future bad interactions, while high-quality context creates a positive one. They link context quality to the 'triple debt' model, which includes technical, cognitive, and intent debt, suggesting that agents make invisible organizational knowledge gaps more visible and costly. The authors argue that organizations must treat context quality as a primary engineering discipline, comparable to code quality, to maximize agent effectiveness and mitigate compounding failures. As Brian Houck noted, the end result of agentic work will be dictated by the quality of the context provided, particularly as agents work more autonomously.