跳转至

文章背景与核心概要

随着编程智能体越来越多地承担复杂的实现任务，核心挑战已从构建单个智能体转变为构建能够系统性提升智能体能力的可扩展基础设施。传统系统往往存在优化孤立、缺乏组合推理能力以及无法根据运行反馈进行演进的问题。

为了解决这些局限性，MEGA（元评估驱动的自适应，Meta Evaluation-Grounded Adaptation） 提出了一种自进化基础设施，其核心包含三层架构：第一层（智慧蒸馏）通过行为聚类和实证 A/B 验证从智能体会话中提取可复用的见解；第二层（类型化智慧图谱）将资产分解为原子的“主旨-上下文-结果”（PCR）单元，利用演绎、溯纳和归纳推理实现超越简单嵌入相似度的组合检索；第三层（协同优化）则通过控制评估编排多智能体工作流，将性能提升归因于特定的策略变更。

最终，MEGA 将智能体优化与知识演进统一为一个内聚的反馈循环，为大模型智能体的持续进化提供了坚实的基础设施支撑。

MEGA: Self-Evolving Agent Optimization Infrastructure via Wisdom Graph

Summary

As coding agents increasingly handle complex implementations, the primary challenge has shifted from building individual agents to creating scalable infrastructure that systematically improves them. Traditional systems suffer from isolated optimizations, lack of compositional reasoning, and an inability to evolve based on operational feedback.

MEGA (Meta Evaluation-Grounded Adaptation) introduces a self-evolving infrastructure that addresses these limitations through a three-layer architecture: 1. Layer 1 (Wisdom Distillation): Extracts reusable insights from agent sessions via behavioral clustering and empirical A/B validation. 2. Layer 2 (Typed Wisdom Graph): Decomposes assets into atomic Primary-Context-Resultant (PCR) units, leveraging deductive, abductive, and inductive reasoning for compositional retrieval beyond simple embedding similarity. 3. Layer 3 (Collaborative Optimization): Orchestrates multi-agent workflows, attributing performance boosts to specific strategy changes via controlled evaluations.

Ultimately, MEGA unifies agent optimization and knowledge evolution into a single, cohesive feedback loop.

Document Metadata

  • arXiv ID: arXiv:2608.10504 [cs.AI]
  • Published: August 11, 2026
  • Authors: Jung Hwan Lee, Kyu Ho Lee, Gwang Hoon Yoo
  • Subjects: Artificial Intelligence (cs.AI)
  • Technical Report: 29 pages, 10 figures

Abstract

随着编程智能体日益承担起代码实现的重任，核心挑战已从构建单个智能体转向构建能够系统性提升智能体能力的基础设施。当前的方法在优化智能体系统时往往无法积累可迁移的知识；即便积累了知识，也缺乏对其进行组合推理的能力；同时，这些知识也缺乏根据运行证据进行自演进的机制。

MEGA（Meta Evaluation-Grounded Adaptation） 作为一种自进化基础设施，填补了这些空白：每个优化周期都会产出持久的资产，对这些资产的组合推理会指导后续的优化，而运行证据则会进一步精炼积累的智慧以及支配该智慧的推理过程。

  • 第 1 层（Layer 1） 通过行为模式聚类和实证 A/B 验证，从智能体会话中提炼出可复用的智慧，将每个处理过程转化为持久资产。
  • 第 2 层（Layer 2） 将这些资产分解为类型化智慧图谱中的原子 PCR（Primary-Context-Resultant，主旨-上下文-结果）单元，并执行演绎、溯纳和归纳推理来扩展隐式关系；然后通过组合检索组装特定于上下文的执行计划，从而浮现出仅靠嵌入相似度无法触及的桥接知识。
  • 第 3 层（Layer 3） 对异构智能体工作流（代码节点、大模型调用和工具使用智能体）执行多智能体协同优化，通过消除数据方差的控制评估，将性能提升效果归因于特定的策略变化。

从第 3 层反馈回来的证据推动了控制智慧组合的策划策略以及跨运行积累的优化轨迹的双向自演进。其结果是一个基础设施：在这里，优化智能体系统与演进指导优化的知识成了一个且唯一的同一过程。

As coding agents increasingly handle implementation, the central challenge shifts from building individual agents to building an infrastructure that systematically improves them. Current approaches optimize agent systems without accumulating transferable knowledge, accumulate knowledge without compositional reasoning over it, and lack a mechanism for that knowledge to self-evolve through operational evidence.

MEGA (Meta Evaluation-Grounded Adaptation) addresses these gaps as a self-evolving infrastructure: each optimization cycle produces durable assets, compositional reasoning over those assets guides subsequent optimization, and operational evidence refines both the accumulated wisdom and the reasoning that governs it.

  • Layer 1 distills reusable wisdom from agent sessions through behavioral-pattern clustering and empirical A/B validation, transforming each process into a durable asset.
  • Layer 2 decomposes these assets into atomic PCR (Primary-Context-Resultant) units within a typed Wisdom Graph and performs deductive, abductive, and inductive reasoning to expand implicit relations; it then assembles context-specific execution plans through compositional retrieval that surfaces bridging knowledge unreachable by embedding similarity alone.
  • Layer 3 performs multi-agent collaborative optimization over heterogeneous agent workflows (code nodes, LLM calls, and tool-using agents), attributing improvement effects to specific strategy changes through controlled evaluation that eliminates data variance.

Evidence fed back from Layer 3 drives the self-evolution of both the curation strategies that govern wisdom composition and the optimization trajectories accumulated across runs. The result is an infrastructure in which optimizing an agent system and evolving the knowledge that guides optimization are one and the same process.

Access & Resources