LangChain on Eval-Driven Development: TDD-style methodology for building AI agents
hwchase17 · x · 2026-11-10
LangChain's blog lays out Eval-Driven Development (EDD), TDD for AI agents: stop tweaking prompts and guessing. The three-step process:
- Define what good looks like: a short spec covering desired behavior, what the agent shouldn't do, and what must keep working.
- Write the eval before you build: establish measurable evals up front.
- Make your change and compare: judge every change via experiment comparisons, since fixing one case may break another.
Key practices: run evals in CI and production, and turn failures you find into new dataset examples, hill-climbing toward the eval. LangChain says it used EDD on its most advanced agents (e.g., IssueBench).
Related event: LangChain Proposes Eval-Driven Development for Building AI Agents(2 posts)→
More from coding & agent
- Dots adds multi-machine control; veteran user asks how it beats his SSH setup — gandamu_ml · 2026-11-10
- simonw highlights Anthropic's trick: tell Claude Fable 5.2 'This codebase is yours' — Kyrannio · 2026-11-10
- 'Claude, build a sandbox startup, make no mistakes' — vibes-coding expectations mocked — tekbog · 2026-11-10
- Grady Booch endorses take that 100% AI-written code is a fad that will walk back — Grady_Booch · 2026-11-10
- Fly.io launches Stripe-built Link Agent Wallet so AI agents can book, order and pay — jeff_weinstein · 2026-11-10
- Fable 5.2 just did PCB routing, the job its poster called the last human engineering task — mallow610 · 2026-11-10