English | 简体中文 | 繁體中文 | 日本語 | 한국어 | Русский | Deutsch | Español
Don't teach the AI new knowledge. Use feedforward, integral control, hierarchical decomposition, and self-monitoring — four concepts — to reshape the agent's memory and skill architecture, transforming it from a passive tool into a self-correcting and evolving collaboration engine.
A complete, reproducible, hands-on tutorial.
It adapts the core ideas of Qian Xuesen's Engineering Cybernetics for modern AI agent platforms (Hermes, Claude with MCP, custom GPTs, etc.). Without adding prompt entries or introducing new models, it uses architectural restructuring to make the agent:
- 🎯 Feedforward for pitfall avoidance — proactively loads your preferences and historical lessons before a task starts
- 🔧 Integral control to prevent oscillation — a one-off complaint only gets a temporary fix; the same feedback twice triggers a systematic change
- 🧹 Hierarchical separation — memory stores only the "why"; skills store only the "how"; never mixed
- 🚦 Self-monitoring delivery gate — a five-check quality gate runs before any output; unqualified results are never delivered
- 🔄 Closed-loop evolution — say "take off" once a week and the agent automatically reviews and optimizes itself
Send this command to your agent, and it will automatically complete the full training (≈ 40 minutes):
按 github.com/zeronesun/cybernetic-your-agent 的 scripts/deployment-commands.md 训练我。每步确认，完成后闭环测试。
Requirements: Your agent must be able to access GitHub, write to memory, and create skills. If not met, follow the manual steps below.
- Make sure your agent platform has: persistent memory + skills / procedural modules
- Read the tutorial (in order is recommended, or at least read Chapter 1 to understand the approach)
- Go to Chapter 5, copy-paste the instructions and deploy step by step (≈ 1 hour)
- Chapter 6 tells you how to use it daily and how to verify it's working
| Chapter | File | One-sentence summary |
|---|---|---|
| Preface | PREFACE.md |
Why AI learns all knowledge but still doesn't "get you" |
| Chapter 1 | chapters/01-background.md |
Background and core philosophy |
| Chapter 2 | chapters/02-prerequisites.md |
Prerequisites and preparation |
| Chapter 3 | chapters/03-core-theory.md |
Four cybernetics concepts in detail (feedforward / integral / separation / self-monitoring) |
| Chapter 4 | chapters/04-architecture.md |
System architecture in detail (three-tier memory + skill layer + data flow) |
| Chapter 5 | chapters/05-deployment.md |
Hands-on guide (step-by-step deployment instructions you can copy directly) |
| Chapter 6 | chapters/06-daily-usage.md |
Daily usage and verification |
| Chapter 7 | chapters/07-faq-extensions.md |
Troubleshooting + multi-agent / new-theory extensions |
| Epilogue | EPILOGUE.md |
After deployment: meaning and future directions |
Root directory and chapter file names use English for cross-platform and toolchain compatibility. Mapping from Chinese source file names to GitHub names:
| Local file name (Chinese) | GitHub file name |
|---|---|
| 序章：为什么 AI…md | PREFACE.md |
| 第一章：背景与核心理念.md | chapters/01-background.md |
| 第二章：适用条件与准备工作.md | chapters/02-prerequisites.md |
| 第三章：核心理论详解.md | chapters/03-core-theory.md |
| 第四章：系统架构详解.md | chapters/04-architecture.md |
| 第五章：实操指南.md | chapters/05-deployment.md |
| 第六章：日常使用与验证.md | chapters/06-daily-usage.md |
| 第七章：常见问题与延伸思考.md | chapters/07-faq-extensions.md |
| 结尾：架构之后.md | EPILOGUE.md |
Headings and body text inside files remain in Chinese. The translations/ directory is reserved for translations into other languages (e.g., en/, ja/).
| File | Description |
|---|---|
scripts/deployment-commands.md |
Pure extraction of all Chapter 5 commands (no explanations, for quick execution) |
examples/ |
L2/L3 example, integral control table example, deep review output example |
CONTRIBUTING.md |
Contribution guide |
LICENSE |
Open source license |
- ✅ Hermes 3 (full native support, all instructions work seamlessly)
- ✅ Claude (with MCP) (minor syntax tweaks for tool calls; core architecture fully applicable)
- ✅ Custom GPTs (use Knowledge files to simulate memory, Actions to simulate skills)
- 🔸 Other agent platforms with persistent memory + skill / procedural modules
- 🔸 Platforms without a Skill feature (can run in a degraded mode, see Chapter 7 §7.1)
| Dimension | Before | After |
|---|---|---|
| Task start | You state the need; agent starts directly | Agent outputs a pitfall checklist and risk assessment first; you confirm |
| Recurring mistakes | Correct every time; same mistake next time | Automatically sinks into rules after the second occurrence; won't happen a third time |
| Memory signal-to-noise | Principles, steps, examples all mixed | Three-layer separation; principles and steps each have their own place |
| Output quality | You inspect, find clichés / formatting issues, then ask for corrections | Agent runs its own five-check quality gate; doesn't deliver until it passes |
| System evolution | Occasionally tweaked based on intuition | Say "take off" weekly; automatic review and auto-optimization |
What you're doing is not prompt engineering. It's architecture engineering.
You're not teaching the agent new knowledge. You're using systems engineering thinking to design an internal architecture that runs stably and evolves continuously.
For more background and elaboration, see the Preface and Chapter 1.
Cybernetic Your Agent — injecting cybernetic genes into your agent.
We don't teach AI new knowledge. We use the four concepts from Qian Xuesen's Engineering Cybernetics — feedforward, integral control, hierarchical decomposition, and self-monitoring — to reshape the agent's memory and skill architecture. The agent moves from understanding commands to understanding your judgment logic — truly becoming your agent.
This project is licensed under the MIT License.
-
Qian Xuesen's Engineering Cybernetics — the theoretical foundation of this tutorial
-
Everyone who helped test and improve the architecture in its early versions