Date: July 4, 2026 | Estimated Reading Time: 31 min | Author: Lilian Weng * Pattern 2: File System as Persistent Memory * Pattern 3: Sub-agent and Backend Jobs
核心贡献 · Key contributions
将“工程化框架”定义为围绕基础模型的系统，协调执行、规划、工具、上下文和评估。 Defines harness engineering as the system around a base model orchestrating execution, planning, tools, context, and evaluation.
提出通过框架进化实现递归自我改进，其中框架代码本身由 LLM 智能体优化。 Proposes recursive self-improvement via harness evolution, where the harness code itself is optimized by LLM agents.
引入元上下文工程（MCE）和元框架作为上下文管理和框架代码的双层优化。 Introduces Meta Context Engineering (MCE) and Meta-Harness as bi-level optimization over context management and harness code.
证明自动化工作流设计（ADAS、AFlow）在编码、问答和数学任务上优于手工设计的工作流。 Demonstrates that automated workflow design (ADAS, AFlow) outperforms handcrafted workflows in coding, QA, and math tasks.
表明框架更新能力在不同模型规模上持平，但框架收益能力在中档模型达到峰值。 Shows that harness updating capability is flat across model sizes, but harness benefit capability peaks at mid-tier models.
强调代码是框架优化的通用语言，能够搜索比提示词更大的设计空间。 Highlights that code is a universal language for harness optimization, enabling search over a larger design space than prompts.
评估通常缓慢、模糊或基于启发式，限制了在具有明确指标的领域的适用性。 Evaluation is often slow, ambiguous, or heuristic-based, limiting applicability to domains with clear metrics.
奖励黑客风险：智能体过度拟合测试、评判模型或基准测试的伪影。 Risk of reward hacking: agents overfit to tests, judge models, or benchmark artifacts.
进化循环中的多样性崩溃：种群收敛到相似的高奖励模式，阻碍开放式发现。 Diversity collapse in evolutionary loops: population converges to similar high-reward patterns, hindering open-ended discovery.
长期成功指标（如可维护性、向后兼容性）无法通过沙盒训练捕获。 Long-term success metrics (e.g., maintainability, backward compatibility) are not captured by sandbox-based training.
模式 2：文件系统作为持久化内存Pattern 2: File System as Persistent Memory#(https://lilianweng.github.io/posts/2026-07-04-harness/#pattern-2-file-system-as-persistent-memory)
模式 3：子智能体与后台任务Pattern 3: Sub-agent and Backend Jobs#(https://lilianweng.github.io/posts/2026-07-04-harness/#pattern-3-sub-agent-and-backend-jobs)