The objective measurement of an AI agent's performance or output against specific, numerically verifiable criteria, rather than subjective assessment. It transforms general notions of "good enough" into a measurable standard against which systems can be reliably compared.
For agentic AI systems performing complex, multi-step tasks, establishing quantifiable quality is crucial for proving value and enabling robust comparison. Without it, claims of cost savings or efficiency are difficult to validate, hindering enterprise adoption and making risk assessment challenging.
A common misconception is that "quality" is inherently subjective or only discernible by human review. While human judgment is often the ultimate arbiter, effective agentic system evaluation requires breaking down complex task outcomes into constituent parts that *can* be objectively measured, such as error rates, adherence to specific rules, or data extraction accuracy.