LLM agents often rely on skills—structured natural-language specifications that determine how they reason, retrieve information, and respond. Unlike monolithic prompts, these skills are multi-field artifacts, meaning they have multiple components that can be tuned.

The paper introduces MOCHA (Multi-Objective Chebyshev Annealing) as a method for optimizing these skills. It frames skill optimization as a multi-objective problem and uses Chebyshev annealing to navigate trade-offs between objectives.

The approach suggests a shift from treating prompts as single blocks to optimizing the structured fields within skills. The abstract does not provide experimental details, but the method is positioned as a way to improve agent behavior through better skill design.