CodeEvolver Blog

We cut a LegalTech company's AI failures by 69%.

CodeEvolver is a coding agent that reflects on the errors in your AI application and changes the prompts and code to improve your system. Here's what it is and why we built it.

Results from running the self-improvement engine on HoVer and PhantomWiki: 80% and 85% relative gains on validation sets.

How to use a training, validation, and testing set to measure overfitting.

AI applications are where self-improving software starts — because that's where the pain is. But the pattern generalizes to any software with a measurable objective.

When ground truth is scarce, an LLM judge can score your outputs — but an optimizer will probe every weakness in that judge. How to build reward functions that survive optimization pressure.

Always keeping the best candidate sounds obviously right. It's a trap: the path to a great system often runs through a temporarily worse one.

An optimizer is only as good as the signal it optimizes against. What goes into an evaluation dataset that won't lead your system astray.

Genetic Pareto optimization is the state of the art in prompt optimization. A plain-English walkthrough of how it searches, why the Pareto frontier matters, and where it stops.

The inspect-tweak-rerun loop feels productive. It isn't. A look at why manual tuning plateaus, and what replacing yourself in the loop actually buys.

Manual prompting is a tedious cycle, and prompt-only optimization has a ceiling. A tour of GEPA, what it achieves, and why the next gains come from evolving the whole system.