How Harness Engineering Turns AI Coding into Real-World Cybernetics

Preface

The author discovered a striking similarity between AI programming and Norbert Wiener’s cybernetics after reading OpenAI’s "Harness engineering: leveraging Codex in an agent‑first world" and George’s analysis "Harness Engineering Is Cybernetics".

Harness Engineering Overview

OpenAI reported that three engineers, over five months, produced 1,500 pull requests, writing one million lines of code entirely generated by AI, claiming a 90% time saving compared with manual coding. The core idea is that humans set the direction while agents execute the code, shifting engineers from writing code to designing environments, intents, and feedback loops.

Key Practices

Keep AGENTS.md short (≈100 lines) as an index rather than a massive instruction file; expose deeper knowledge progressively.

Store all rules, architectural decisions, and conventions in a version‑controlled repository so the AI can read them.

Enforce architectural constraints (e.g., three‑layer structure: gateway, domain, DAO) through scripts or pipelines that validate generated code.

Build an observable AI system where logs, monitoring, and test results are fed back to the agent for self‑correction.

Automate cleanup of "garbage" code by periodically scanning the repository and generating refactoring pull requests.

Cybernetics Basics

Cybernetics studies how systems use information to achieve purposeful behavior. It focuses on three core concepts: information, control, and feedback.

Information

Information reduces uncertainty. For example, a colleague telling you that an article is "great" eliminates doubt about its quality. In cybernetics, information is the exchange between a system and its environment that enables adaptation.

Control

Control adjusts a system’s state to reach a target. The author illustrates this with a personal relationship example, showing how multiple possible outcomes can be steered toward a desired one.

Feedback

Feedback returns the system’s output to the input for comparison with the goal, allowing self‑correction. Negative feedback expands control capability, as seen in physiological regulation (e.g., insulin lowering blood sugar).

Mapping to AI Programming

The control loop maps to AI‑driven development as follows:

Goal: The desired feature or behavior.

Controller: The AI agent.

Controlled System: The codebase.

Sensors: QA tools, static analysis, test suites, and deployment logs that detect deviations.

By repeatedly feeding sensor data back to the agent, the AI can iteratively narrow the "possibility space" of code implementations, similar to a multi‑step control process.

Industry Perspectives

George highlighted three real‑world examples of the same pattern:

A centrifugal governor that senses speed and adjusts a valve.

Kubernetes controllers that reconcile desired and actual cluster state.

OpenAI’s Harness Engineering, where engineers design the environment and let agents write, test, and merge code.

Personal Viewpoint

The author argues that Harness Engineering is a concrete implementation of cybernetics in software. By encoding business rules, architectural constraints, and quality standards into machine‑readable artifacts, developers turn AI from a fast code generator into a reliable engineering partner.

Implications for Developers

Programmers must shift from "writing code" to "designing the environment". Their core value becomes evaluation: defining correct behavior, setting intents, and building feedback mechanisms. This mirrors the historical transition from manual valve operators to control‑system designers.

Conclusion

In the AI era, the programmer’s primary role is to create and maintain the rule set, feedback loops, and observability infrastructure that guide agents. Success depends on robust tooling—syntax checkers, security scanners, unit and integration tests—rather than raw coding speed.

Illustrative Example

12个小球，能否用天平称3次，找出唯一的、轻重未知的那个小球？
这是一个数学问题，我记得某次数学竞赛中做过，当时费了好大劲来得到答案，那么我们如果用控制能力来分析，我们来看下多简单：
未称之前，每个小球都可能是废品，废品有轻重两个状态，可能性空间为12*2=24
天平每称一次控制能力为3（左边重、右边重、一样重），即可能性空间缩小为原来的1/3
称三次控制能力为3*3*3=27，27>24，因此可以解决问题。

Key Diagrams