# Agentic Context Engineering Agentic Context Engineering (ACE) is a framework where AI agents actively manage and evolve their own context rather than relying on humans to curate it. Unlike traditional [[Context Engineering]] where humans design the context, ACE treats contexts as evolving playbooks that accumulate, refine, and organize strategies through a modular process of generation, reflection, and curation. The key shift: in standard context engineering, humans decide what context to include. In ACE, agents participate in that decision. They reflect on what worked, what didn't, and incrementally update their own context to improve future performance. ## Core mechanisms 1. **Generation**: agents produce outputs using their current context 2. **Reflection**: agents evaluate their own outputs against success criteria 3. **Curation**: agents update their context based on what they learned, adding strategies that worked and pruning what didn't 4. **Structured updates**: incremental, modular changes that preserve detailed knowledge rather than wholesale rewrites This prevents context collapse; the problem where aggressive context compression or rewriting destroys nuanced knowledge that took many iterations to build. ## Connection to PCM and harness engineering ACE is what happens when [[Personal Context Management (PCM)]] becomes bidirectional: not just you curating context for AI, but AI curating its own context based on experience. In a well-designed [[Harness Engineering|harness]], ACE operates within human-defined boundaries while still having autonomy over its own learning. This is already happening in practice with AI agent memory systems that accumulate lessons, feedback, and patterns across conversations. ## References - https://arxiv.org/abs/2510.04618 ## Related - [[Context Engineering]] - [[Personal Context Management (PCM)]] - [[Harness Engineering]] - [[AI Agent Memory]] - [[AI Agent Skills]] - [[Context Drift]] - [[Context Hygiene]] - [[AI context is finite with diminishing returns]] - [[Levels of AI Context Management]]