Agent Kung-fu: 10 Layers of Control Transfer for Production-Ready Agents
This article presents a 10-level framework for building production-ready AI agents, centered on six critical questions about action control, capability access, conflict resolution, failure handling, knowledge reuse, and causal reasoning, plus four engineering principles emphasizing code-enforced permissions, evaluator-driven loops, identification before optimization, and falsifiable innovation.
Six Questions That Determine Production Readiness
The author argues that whether an agent can reach production depends not on a slightly stronger model, but on having answers to six fundamental questions:
Who decides the next action? — Control over the decision loop.
What capabilities can the model call? — Tool and skill boundaries.
When multiple participants disagree, who arbitrates? — Conflict resolution mechanism.
How does failure enter the next round instead of being silently ignored? — Explicit failure handling and retry logic.
How does experience precipitate into reusable knowledge? — Knowledge accumulation and reuse.
When can we only describe correlation, and when can we talk about causal intervention? — Distinction between observational and interventional reasoning.
Ten Layers, Ten Transfers of Control
The framework consists of ten layers; each layer transfers one more piece of control from the model to code, protocols, governance, feedback, data relationships, knowledge, semantics, causality, and finally to falsifiable experiments. The layers are not merely renamed chat scripts — they represent a progressive handoff of authority.
One Persistent Task: Top-5 Movie Categories by Revenue
The entire book uses a single, unchanging task: "Find the top 5 movie categories by revenue." What changes across layers is who controls the process. In Level 1, code handles fixed reading, connection, and aggregation; the model only submits a number N.
Four Core Principles (心法)
Prompt is a soft constraint. Permissions must be enforced at the code, sandbox, or database layer.
A loop without an evaluator is mere repetition; a loop without a stop condition is a risk amplifier.
Identify first, then optimize. Observing a difference does not automatically become an intervention effect.
Innovation must be falsifiable. A mechanism that cannot explain how it fails is not a testable solution.
How to Use This Material
90-Minute Quick Course
Run Levels 1–3 to observe how answer, tool, and trace change; then run Levels 4–5 to read evaluation feedback; finally rewrite the "conclusions" of Levels 9–10 into "identify gaps + experiments."
One-Day Engineering Workshop
Add a read-only account and timeout to the SQL guard; modify Level 4's action order, N, and skill summary; verify that all three cases are intercepted before execution; finally write a pre-registered inventory experiment plan.
Ten layers are not a leaderboard. Choose the lowest layer that is sufficient for your control needs.
Resources
Book PDF:
https://raw.githubusercontent.com/bibiparrot/agent_kungfu_levels/main/agent-kungfu-ten-levels-zh.pdfCode repository:
https://github.com/bibiparrot/agent_kungfu_levelsSigned-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
AI2ML AI to Machine Learning
Original articles on artificial intelligence and machine learning, deep optimization. Less is more, life is simple! Shi Chunqi
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.