Agentic Context Engineering: Evolving Contexts for SelfImproving Language Models (arxiv.org)

Researchers introduced ACE (Agentic Context Engineering), a framework that treats prompts, memories, and other context artifacts as evolving “playbooks” that are generated, reflected on, and curated over time instead of being replaced or compressed. ACE builds on Dynamic Cheatsheet’s adaptive memory and directly tackles two common failures of context-based adaptation: brevity bias (loss of domain nuance when compressing information) and context collapse (iterative rewrites that erode detail). By applying modular, incremental updates that preserve and organize granular knowledge, ACE works across both offline settings (system prompts, domain playbooks) and online agent memory, and is designed to leverage long-context LLMs to avoid information loss. Empirically, ACE improves performance and efficiency: +10.6% on agent benchmarks and +8.6% on finance tasks versus strong baselines, lowers adaptation latency and rollout cost, and can learn from natural execution feedback without labeled supervision. On the AppWorld leaderboard ACE matches the top production agent on average and outperforms it on the hardest test split while using a smaller open-source model — demonstrating that evolving, structured contexts enable scalable, self‑improving LLM systems. For the AI/ML community this offers a practical, low‑overhead alternative to weight updates for continual adaptation, especially for agent frameworks and domain-specific reasoning that must retain rich, long-lived knowledge.