Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Meta Context Engineering via Agentic Skill Evolution

Haoran Ye, Xuning He, Vincent Arak, Haonan Dong, Guojie Song

TL;DR

Meta Context Engineering (MCE) reframes context engineering as a bi-level, learnable optimization where CE skills and context artifacts co-evolve. A meta-agent evolves CE skills via agentic crossover, while a base-agent implements those skills to optimize context as flexible files and code, enabling rapid adaptation and self-improvement of LLM-driven systems. Across five domains and multiple models, MCE yields consistent gains over state-of-the-art CE methods, with pronounced benefits for smaller models, and demonstrates superior context adaptability, transferability, and training efficiency. The work introduces a scalable design space for agentic AI, supported by extensive ablations and analyses, and outlines future directions toward broader, autonomous skill evolution and more efficient interaction with context artifacts.

Abstract

The operational efficacy of large language models relies heavily on their inference-time context. This has established Context Engineering (CE) as a formal discipline for optimizing these inputs. Current CE methods rely on manually crafted harnesses, such as rigid generation-reflection workflows and predefined context schemas. They impose structural biases and restrict context optimization to a narrow, intuition-bound design space. To address this, we introduce Meta Context Engineering (MCE), a bi-level framework that supersedes static CE heuristics by co-evolving CE skills and context artifacts. In MCE iterations, a meta-level agent refines engineering skills via agentic crossover, a deliberative search over the history of skills, their executions, and evaluations. A base-level agent executes these skills, learns from training rollouts, and optimizes context as flexible files and code. We evaluate MCE across five disparate domains under offline and online settings. MCE demonstrates consistent performance gains, achieving 5.6--53.8% relative improvement over state-of-the-art agentic CE methods (mean of 16.9%), while maintaining superior context adaptability, transferability, and efficiency in both context usage and training.

Meta Context Engineering via Agentic Skill Evolution

TL;DR

Meta Context Engineering (MCE) reframes context engineering as a bi-level, learnable optimization where CE skills and context artifacts co-evolve. A meta-agent evolves CE skills via agentic crossover, while a base-agent implements those skills to optimize context as flexible files and code, enabling rapid adaptation and self-improvement of LLM-driven systems. Across five domains and multiple models, MCE yields consistent gains over state-of-the-art CE methods, with pronounced benefits for smaller models, and demonstrates superior context adaptability, transferability, and training efficiency. The work introduces a scalable design space for agentic AI, supported by extensive ablations and analyses, and outlines future directions toward broader, autonomous skill evolution and more efficient interaction with context artifacts.

Abstract

The operational efficacy of large language models relies heavily on their inference-time context. This has established Context Engineering (CE) as a formal discipline for optimizing these inputs. Current CE methods rely on manually crafted harnesses, such as rigid generation-reflection workflows and predefined context schemas. They impose structural biases and restrict context optimization to a narrow, intuition-bound design space. To address this, we introduce Meta Context Engineering (MCE), a bi-level framework that supersedes static CE heuristics by co-evolving CE skills and context artifacts. In MCE iterations, a meta-level agent refines engineering skills via agentic crossover, a deliberative search over the history of skills, their executions, and evaluations. A base-level agent executes these skills, learns from training rollouts, and optimizes context as flexible files and code. We evaluate MCE across five disparate domains under offline and online settings. MCE demonstrates consistent performance gains, achieving 5.6--53.8% relative improvement over state-of-the-art agentic CE methods (mean of 16.9%), while maintaining superior context adaptability, transferability, and efficiency in both context usage and training.
Paper Structure (54 sections, 5 equations, 4 figures, 4 tables, 1 algorithm)

This paper contains 54 sections, 5 equations, 4 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Agentic Context Engineering
  4. Evolutionary Computation with LLMs
  5. Meta Context Engineering
  6. Problem Formulation
  7. Meta-Level: Agentic Skill Evolution
  8. Base-Level: Fully Agentic Context Optimization
  9. Algorithmic Orchestration
  10. Experiments
  11. Experimental Setup
  12. Tasks and Datasets.
  13. Baselines.
  14. Models.
  15. MCE Instantiation.
  16. ...and 39 more sections

Figures (4)

  • Figure 1: Conceptual overview of Meta Context Engineering (MCE). A cartoon representation of the bi-level optimization framework. The meta-agent drives skill evolution while the base-agent manages context optimization. This dual-layered approach ensures the agentic harness and context artifacts co-evolve for maximum performance.
  • Figure 2: Methodological overview of Meta Context Engineering (MCE).
  • Figure 3: Context efficiency on FiNER. We plot context accuracy vs. context tokens. For MCE, we include the context generated in two iterations. For ACE, we include context at Step 20 of Epoch 1, end of Epoch 1, 2, and 5.
  • Figure 4: Training efficiency on FiNER. We plot best-so-far training set accuracy vs. the number of rollouts (for MCE, this includes both training and validation inference). We also indicate the total training duration.