Agentic Context Engineering (ACE) is a technique jointly proposed by a research team from Stanford University, SambaNova, and UC Berkeley in October 2025. The related research results were published in the paper "[…]".Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models".

Agentic Context Engineering is a framework that views context as an evolutionary script accumulated, refined, and organized through a modular process of generation, reflection, and curation. It prevents crashes through structured, incremental updates while preserving detailed knowledge and scaling to long context models. This allows for comprehensive context adaptation in both offline environments (e.g., system hint optimization) and online environments (e.g., memory adaptation during testing). On the AppWorld leaderboard, ACE, using a smaller, open-source model, achieves overall average performance comparable to top-tier production agents.