Insights·2026-06-10

What Is Harness Engineering?

Harness engineering is the practice of improving AI output quality not by changing the model itself, but by designing which tools the model can use and which rules and workflows it must operate within. The same model performs very differently depending on its harness — so before upgrading to a more expensive model, this is the layer to inspect first.

Why the Harness Comes After Prompt Engineering

In the Q&A era of AI, the playbook was to polish prompts and feed context carefully through RAG. Models were weak, so inputs had to be packaged with care. Today's models are different: smart and broadly capable, but prone to running off in the wrong direction when left unconstrained.

The focus has therefore shifted from packaging inputs to designing behavior — deciding which tools an agent may use (file read/write, web search, terminal), in what order it works, and what it must never do. That is harness engineering, and it is becoming the next stage after prompt engineering.

Same Model, Different Results — What the Harness Decides

Notably, performance changes without changing the model. One open-source agent measurably improved coding performance simply by changing the format in which the AI outputs code edits. How an agent locates the edit site and verifies results matters as much as the model's raw intelligence.

This has cost implications. Switching to a pricier model is the easy option, but tuning the harness is often cheaper and more effective. Much of "I tried AI and it wasn't great" reflects a missing harness, not a model limitation.

Harness Engineering Without Writing Code

Modifying agent code requires engineering skills, but the entry point is far lower. Writing rules into markdown files like AGENTS.md or SKILL.md — "never do this," "the project is structured like that," "output in this format" — is harness engineering too. Agents read these documents before working and follow them.

It resembles dog training: the training method matters more than acquiring a better dog. And as with dogs, long convoluted commands stop landing — keep rules short, clear, and split into small documents.

The AX View: Document the Rules Before Buying Tools

This is why SH Consulting gives harness engineering significant weight in AX (AI Transformation) training. What an organization needs first is not another tool, but the habit of translating its rules, constraints, and judgment criteria into documents AI can read. Once those documents accumulate, any agent you attach will work the way your organization works.

Source: 유튜브 '요즘 판교사투리 1. 하네스 엔지니어링' 영상을 보고 정리