Self-improving agents: continuous evals, RL loops and the production optimization moat

Frontier models are converging.

The optimization stack

1. 1.Continuous evals.
2. 2.Trace mining.
3. 3.Reward modeling.

Why most teams stop at evals

"Once the loop closed, our agent quality improved every week without anyone explicitly working on it."

What 'self-improving' is not

Operating discipline

• Version everything.
• Reward-model conflicts must surface, not silently overwrite.
• Every shipped change has a rollback plan and a measured outcome.