meta-context-engineering
Meta Context Engineering via Agentic Skill Evolution is a bi-level agentic framework designed to optimize context for large language models by replacing rigid, manually crafted heuristics with learnable skills. This system co-evolves context engineering skills and context artifacts, treating the optimization methodology itself as a learnable object. Unlike traditional approaches restricted to fixed workflows like prompt rewriting or additive curation, MCE employs a two-tier optimization process. At the meta-level, agents analyze task specifications and performance history to generate improved skills through agentic crossover, including executable code and dynamic operators. At the base level, the system executes these skills without structural constraints to produce optimized context functions comprising both static components and dynamic operators. MCE demonstrates significant performance gains across diverse domains including finance, chemistry, medicine, law, and AI safety, achieving roughly 89 percent rel