The Paradigm Shift: Output vs. Behavior

Traditional LLM Evaluation

Agent Evaluation

Goal
Environment
Execution
Failure Mode

Output Accuracy
Static Datasets
Single-path Processing
Hallucination

Workflow Behavior
Dynamic Contexts
Multi-path & Tool Dependent
Cascading Workflow Failure