Use Quality Dimensions, Not One Score
Separate agent quality into dimensions that can be evaluated independently.
The danger is not a low score. The danger is a high average that hides a launch-blocking dimension. Functional fit Functional fit asks whether the agent completes the job it exists to do. In agent evaluation, this includes choosing the right tool, using the right evidence, following the workflow, and producing an output that satisfies the user's goal. Reliability and recovery Reliability asks whether the agent behaves consistently across messy real conditions. Agents fail here when they cannot recover from missing files, unavailable tools, partial context, long threads, or ambiguous instructions. Security and usability Security asks whether the agent protects…
Sign up free — one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in