Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models

Abstract

Agentic Context Engineering (ACE) is a framework that optimizes contexts both offline (e.g., system prompts) and online (e.g., agent memory). ACE treats contexts as evolving playbooks that accumulate, refine, and organize strategies through a modular process of generation, reflection, and curation. Unlike methods that collapse prompts into short summaries and lose detail, ACE uses structured incremental updates that preserve knowledge and scale with long-context models. ACE consistently outperforms strong baselines: +10.6% on agent benchmarks and +8.6% on finance tasks, while significantly reducing adaptation latency and rollout cost.

Relevance to Tokalator

ACE provides theoretical grounding for Tokalator's optimization-first design philosophy: proactive context management yields measurable quality gains (+10.6%) compared to passive acceptance of whatever enters the context window. Tokalator's tab relevance scorer and /optimize command implement the context selection layer that ACE assumes as given.