Agentic Context Engineering (ACE) is a self-improvement method in which an agent curates its own context as a growing “playbook.” Rather than updating the model’s weights, the agent accumulates and refines the information it carries in context, so it improves from experience over time.

How it works

ACE builds on /glossary/context-engineering/ and organizes self-improvement around three roles. A Generator produces candidate behaviour and outputs. A Reflector reviews that behaviour and extracts lessons. A Curator decides what to keep, folding useful detail into the evolving playbook. This treats context as durable /glossary/agent-memory/ that the agent maintains itself.

Because the playbook grows through curation, ACE lets an agent improve without /glossary/fine-tuning/, avoiding the cost of retraining. The curation step is designed to prevent “context collapse,” the loss of accumulated detail that can occur when context is rewritten carelessly, a failure mode related to /glossary/context-rot/.

Origins and History

Agentic Context Engineering was introduced by Zhang, Hu, Olukotun, Zou, and colleagues at Stanford, SambaNova, and UC Berkeley in “Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models” (arXiv 2510.04618), released on 6 October 2025. The paper defines the Generator, Reflector, and Curator roles and frames ACE as a way for language models to self-improve by evolving their contexts while avoiding context collapse.

Sources

  1. Zhang, Hu, Olukotun, Zou, et al. (Stanford, SambaNova, UC Berkeley). “Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models” (6 October 2025). https://arxiv.org/abs/2510.04618