- AI agents are non-deterministic, meaning the same input can produce different outputs, sometimes with negative consequences.
- Evaluations are essential for monitoring AI system performance, involving two main patterns: offline and online evals.
- Offline evals use a fixed dataset before deployment, like unit tests, to catch regressions in controlled cases.
- Online evals score each interaction in real-time against live production data, providing a more accurate signal of actual behavior.
- Good evals consist of a dataset, split testing/experiments, and scoring methods such as LLM-as-Judge, algorithmic, and signal-based.
- Offline evals offer pre-deployment peace of mind but are limited to predefined cases and require ongoing dataset maintenance.
- Online evals leverage real data, support split-test experiments, and can use deferred scoring for delayed outcomes, but lack pre-deploy gates.
- Using both offline and online evals is recommended to cover different risks: offline for pre-release checks and online for real-world performance.