Loop Engineering vs. Harness Engineering

Since the rise of AI agents, the industry has been talking about:

• Prompt Engineering
• Context Engineering
• Loop Engineering
• Multi-Agent Systems But there's another concept that has quietly become increasingly important:

Harness Engineering.

Many people see Claude Code, Codex, or Cursor automatically modifying code and conclude: The future is simply letting agents run in loops by themselves. However, anyone who has worked on large-scale projects quickly discovers something else: Very often, the problem isn't that AI can't work. The problem is: AI doesn't have a sufficiently good working environment. That's why Harness Engineering exists.

The Short Answer

Loop and Harness are not competing ideas. They solve two entirely different problems.

Harness Engineering↓Enables AI to work.
Loop Engineering↓Enables AI to keep working.

Or, put another way:

Harness:What can the AI do?
Loop:How far can the AI go by itself?

What Is Harness Engineering?

The word "harness" originally means:

• A safety harness
• A horse harness
• A control system

In AI, it can be understood as: Providing AI with a complete, operable, and observable working environment. Examples include:

• Terminal
• Git
• Browser
• Database
• Documentation
• Screenshots
• Figma
• Playwright
• Design Systems
• MCP Servers

With a good harness, AI doesn't just answer questions.

It can actually:

• Read code
• Modify files
• Run tests
• View applications
• Operate a browser
• Create pull requests

All of these capabilities together form the Harness.

What Is Loop Engineering?

Loop Engineering is fundamentally different.

It focuses on: How AI can continuously move work forward until a goal is achieved. For example:

PLAN↓EXECUTE↓VERIFY↓REPLAN↓EXECUTE

Until: Success Or: A reasonable stopping point. The core components of a Loop usually include:

• Goal
• State
• Verification
• Feedback
• Stop Conditions

A Layered View of AI Development

I like to think about AI engineering in the following layers:

Level 1Prompt Engineering
Level 2Context Engineering
Level 3Harness Engineering
Level 4Loop Engineering
Level 5Multi-Agent Systems

Each layer builds upon the previous one. Without a good Harness, it's difficult for a Loop to realize its full potential.

Harness Solves Discovery Problems, Loop Solves Optimization Problems - This is the biggest difference between the two.

A Real-World Example

Imagine a team is building a web application. The team doesn't have a UI/UX designer. They only have a vague requirement:

We want to build a dashboard.

• A sidebar on the left
• A navbar at the top
• KPI cards in the center
• Recent orders at the bottom

But they don't have:

• Figma designs
• A complete PRD
• Detailed user flows
• Loading states
• Error states
• Empty states
• Mobile designs

This situation is actually very common.

What Happens If You Use a Loop Immediately?

You might define the goal as:

Goal:Build the Dashboard page.

But then a problem arises.

What does "done" actually mean?

• How large should the KPI cards be?
• Should the table have pagination?
• How should filtering work?
• How wide should the sidebar be?

The answers don't exist yet. There is no objective definition of success. And therefore, there is no: Verifier Without a verifier, a Loop can easily:

• Keep modifying
• Keep retrying
• Keep consuming money While never actually knowing what success looks like.

This Is Where Harness Shines What we really need is:

Markdown↓AI generates a page↓Preview in a browser↓Take screenshots↓Human provides feedback↓AI modifies the page↓Generate again

For example:

# Dashboard
Sidebar on the left.
Navbar at the top.
Display four KPI cards.
Display the latest ten orders.

AI: Generate the React page.

You:

The KPI cards are too large.The sidebar is too wide.The table isn't clear enough.

AI:

Modify.Regenerate.

This is a: Human-in-the-loop discovery process.

This Is Not a Loop This is a: Harness Workflow Because the requirements themselves are still evolving.

When Should You Use Harness?

Harness delivers the most value when:

• Requirements are ambiguous.
• The product is still being explored.
• There are no finalized designs.
• Human judgment is heavily involved.
• The UI changes frequently.
• Success criteria are unclear.

When Should You Use Loops? Loops shine when:

• Requirements are stable.
• Objective verification exists.
• Tests are available.
• A Design System exists.
• Success criteria are clearly defined.

For example:

• Fix lint errors
• Fix failing tests
• Improve accessibility
• Fix responsive issues
• Perform automated code reviews

In these scenarios, Loops can provide enormous productivity gains.

Web Development Typically Goes Through Three Stages

Stage 1: Discovery

• Requirements are vague.
• The UI is undefined.
• The product is still taking shape.

Most important: Harness Engineering

Stage 2: Implementation

Requirements gradually become stable.

Most important: Harness Engineering + Small Loops

Stage 3: Maintenance

Requirements are stable. Most important: Loop Engineering

One-Sentence Summary

Harness Engineering determines what AI can do.Loop Engineering determines how much AI can accomplish on its own.

Or:

Harness is the working environment.Loop is the workflow.

Or even:

Harness allows AI to see, touch, and operate.Loop allows AI to keep moving forward.

They are not substitutes.

They are complementary.

However, during the early stages of product and web development—especially when:

• There is no UI/UX designer
• Websites are defined through Markdown or natural language
• Requirements change frequently
• Humans are still discovering the product

Harness Engineering often matters far more than Loop Engineering.

Because:

Before you know where you're going, moving faster is less important than seeing more clearly.