xiaomo
04e2b5b8a7
CI / check (pull_request) Successful in 2m21s
26 cards covering: - Project philosophy (session isolation, process discipline, dissipative structure) - Comparisons (vs skill, vs dynamic workflow) - Business rationale (FTE vs vendor, OPC, switching cost) - Learning model (memory + skill + workflow) - Self-improvement (reflective workflow, eval) - Open questions (workflow granularity, human-in-the-loop)
36 lines
1.4 KiB
Markdown
36 lines
1.4 KiB
Markdown
---
|
|
title: "Reflective Workflow — Self-Improvement as Discipline"
|
|
created: "2026-06-07"
|
|
source: "openclaw-xiaomo"
|
|
tags: [architecture, pattern, decision]
|
|
category: "architecture"
|
|
links:
|
|
- eval-closes-the-trust-chain
|
|
- three-learning-carriers
|
|
- workflow-as-improvable-system
|
|
- feedback-loops-convergent-and-divergent
|
|
- trust-chain-audit-evaluate-reuse
|
|
---
|
|
|
|
FTE agent 的"成长"不靠自发顿悟，靠纪律性的反思。反思本身是纪律性的（定期跑、不能跳过、有固定步骤），所以应该用 workflow 承载——不能靠 agent "有空想想"。
|
|
|
|
反思 workflow 定期拉取最近执行过的任务，分析流程中出现的问题，找可优化的点，迭代，eval，对比。反思的对象覆盖三层载体：
|
|
|
|
- 发现某个 role 反复在同一类问题上出错 → **迭代 skill**
|
|
- 发现某类任务的上下文总是缺少关键信息 → **补充记忆**
|
|
- 发现某个审批环节通过率 100% 从未驳回 → **简化 workflow**
|
|
|
|
这形成了双层 workflow 架构：
|
|
|
|
```
|
|
执行层：workflow 驱动日常任务
|
|
↓ 产出执行记录（CAS 链）
|
|
反思层：反思 workflow 定期分析执行记录
|
|
↓ 产出改进建议
|
|
改进层：迭代 memory / skill / workflow
|
|
↓ 提升下一轮执行质量
|
|
执行层：...
|
|
```
|
|
|
|
两层都是 workflow，职责不同——执行层做事，反思层改进做事的方式。用 workflow 来优化 workflow——工具改进自身的递归。
|