Eval-driven development: write the eval before you build the agent
The point
LangChain defines eval-driven development in three steps: define what good looks like, write the eval before you build, then change the agent and compare. It borrows from test-driven development, but agents need flexible scoring because outputs vary and more than one answer can be right.
Why it matters
For builders, writing the eval first makes progress measurable instead of guessed from hand tests.
Key details
- Score with code evals for deterministic checks and LLM-as-a-judge for the rest
- Run a baseline experiment before any change, then compare each new experiment against it
Original ~8 min → 25 sec here Source: LangChain Blog, Nov 9, 2026