Eval Harness, Evaluation-Driven Development

Evals are to agent development what tests are to traditional software. This skill provides a structured workflow for defining, running, grading, and tracking evaluations against agent features, LLM outputs, or any system where correctness is probabilistic rather than binary.

When to Activate

• Building or modifying agent capabilities
• Validating LLM prompt changes
• Checking for regressions after code changes
• Establishing quality baselines for new features
• Comparing model or prompt variants
• Any time "does this work?" requires more than a unit test

Core Concepts

1. Eval Types