Evals are the gradient signal for harness engineering — the same data quality rigor from ML training applies

The mapping is direct: model + training data + gradient descent → better model, and harness + evals + harness engineering → better agent. Each eval case contributes a signal — “did the agent take the right action?” — that guides the next proposed edit to the harness. This means the same rigor around data quality and curation that determines model training outcomes also determines A mediocre agent inside a strong harness outperforms a stronger agent inside a messy one.

This reframes how to invest in agent improvement. If Evals are behavioral pressure vectors, not neutral measurements — poorly chosen evals distort agent development, then curating those pressure vectors with the same care as ML training data is the highest-leverage activity. A small set of well-tagged evals covering the behaviors you care about beats thousands of noisy high-coverage evals — quality over quantity, exactly like training data. The implication for Agents learn at three distinct layers — model weights, harness code, and context configuration is that the harness layer has its own training loop, distinct from model fine-tuning but equally rigorous.