工程脚手架

Harness Engineering

同一个 AI 有时挺好用，换个地方又犯同样的错。除了模型，还能改什么？

evals

✓ Test 1: pass
✓ Test 2: pass
✗ Test 3: fail
  → improve harness

iterate prompts + tools · measure

harness-engineering · 图文示意