🎉 Giảm 20% cho đơn hàng đầu tiên!✦🚚 Miễn phí vận chuyển cho đơn trên 500.000₫✦⚡ Flash Sale — Số lượng có hạn!✦🎁 Quà tặng kèm cho đơn từ 1.000.000₫✦🎉 Giảm 20% cho đơn hàng đầu tiên!✦🚚 Miễn phí vận chuyển cho đơn trên 500.000₫✦⚡ Flash Sale — Số lượng có hạn!✦🎁 Quà tặng kèm cho đơn từ 1.000.000₫✦
The End of Harness Engineering: when AI writes and grades its own harness, humans are left with the scoring set
Quay lại Blog

The End of Harness Engineering: when AI writes and grades its own harness, humans are left with the scoring set

16 tháng 11, 20268 phút đọcVũ Khôi
English
#English
Vũ Khôi

Vũ Khôi

Tác giả tại Wotech Electronics Blog

Summary image of the paper The End of Harness Engineering

The End of Harness Engineering: when AI writes and grades its own harness, humans are left with the scoring set

Chinese researchers have just published a paper with a title alarming enough to make every AI agent framework nervous: "The End of Harness Engineering". According to a post by the page Học lập trình cho người mới bắt đầu on 15 November 2026, the paper's argument can be summed up in one sentence: humans hand-writing the programs that wrap around AI models is gradually becoming an obsolete skill, because an AI loop can generate, evaluate, and improve the harness itself — leaving humans with only one remaining role: designing the scoring set.

Quick answer: According to the post, the paper "The End of Harness Engineering" argues that a human-written harness is only a guess about how to solve a specific task at a specific moment, deployed and then left running for months without being re-read or evaluated. In contrast, a harness generated by an AI loop is a file capable of evaluating itself: a small 4B model reads the errors that occurred and proposes 8 edits, only the edit scoring highest on an independent test question set is kept, and the large model performing the main task is kept completely unchanged. The human role narrows to a single component: the scoring set — the collection of questions used to grade the harness. Frameworks become continuously replaceable and disposable, while designing the test becomes the core job.

A harness is the program wrapped around an AI model: the code handling prompts, tool calls, think–act loops, result checking, and error handling — in short, following the formula Wotech previously introduced: Agent = Model + Harness. A scoring set is a collection of independent test questions or tasks used to grade and select between harness versions.

What does "The End of Harness Engineering" argue?

According to the post, the Chinese researchers argue that humans hand-writing the programs wrapped around AI models is gradually becoming an obsolete skill. A human-written harness is, at its core, only a guess about how to solve a specific task at a specific moment. It gets deployed, then runs for months without being re-read or evaluated — meaning the initial guess simply persists, even as the task, the data, or the underlying model changes.

A harness generated by an AI loop, by contrast, is a file capable of evaluating itself. Instead of a static human guess, the system continuously measures, proposes changes, and keeps only the changes that actually improve the score. The entire paper, according to the post, is summarized in a single image shared by Mr Buzz — also the only illustration in the original post.

Note: this article presents the paper's argument as described in the Facebook post; Wotech has not read or independently verified the full text of the study.

How does the AI loop that generates a harness work?

As described in the post, the loop works as follows:

  1. A small model reads the errors: A small 4B model (4 billion parameters) reads the errors that occurred while the harness ran.
  2. It proposes several edits: The small model proposes 8 edits to the harness.
  3. Independent scoring: Each edit is scored on an independent set of test questions — separate from the data used to generate the edits.
  4. Only the best is kept: Only the highest-scoring edit is kept; the rest are discarded.
  5. The large model stays untouched: The large model performing the main task is kept completely unchanged, with no weight updates throughout the process.

The post gives an example: place a large language model in the role of performing the main task, and with each loop iteration, the system rewrites and optimizes the harness around it. The harness is therefore no longer a static artefact written once by a human, but a living file, continuously rewritten based on measured evidence.

How does a human-written harness differ from an AI-loop-generated one?

Aspect Human-written harness AI-loop-generated harness
Nature A guess about how to solve a task at one moment A file capable of evaluating itself
Effectiveness evaluation Usually never re-read or evaluated after deployment Scored every iteration on an independent test set
Improvement Depends on human time and attention Proposes several edits itself, keeps only the highest-scoring one
Main model May be changed together with the harness Kept completely unchanged, no weight updates
Lifespan Runs unchanged for months Continuously rewritten and optimized each iteration

What human role remains?

The central question the post raises is: what is left for humans? The paper's answer, according to the post, is that the role narrows to a single component: the scoring set — the collection of questions used to grade the harness.

The implication is striking. Frameworks become continuously replaceable and disposable: if the harness can rewrite and grade itself, value no longer lies in any particular framework, but in the quality of the test used to choose between harnesses. Designing the test — picking the right questions, tasks, and grading criteria — becomes the core job.

In other words, humans move from writing the "answer" (the harness) to writing the "exam" (the scoring set). A poor scoring set will select a poor harness, no matter how many times the loop runs; a good scoring set is what decides how good the final system becomes.

What does this mean for people building AI agents?

If the paper's argument holds, several shifts are worth watching:

  • Hand-writing harnesses loses value: Manually writing prompts, loops, and error handling around a model may become an obsolete skill, as the paper claims.
  • Evaluation design takes center stage: Designing independent test sets, grading criteria, and selection procedures becomes the core work of agent builders.
  • Frameworks become disposable: Rather than committing long-term to one framework, teams may continuously replace and discard frameworks as the loop finds better harnesses.
  • The large model stands still while the harness moves: Keeping the main model unchanged while optimizing the harness shows that performance gains can come from the surrounding layer, without necessarily swapping models or fine-tuning weights.

Caution is still warranted: this is one paper's argument, summarized through a Facebook post and a single image. The method's scope — which task types it suits, what the loop costs to run, how large a scoring set must be — is not presented in the post, and Wotech has not independently verified these claims.

Frequently asked questions

1. What is "The End of Harness Engineering"? The title of a paper by Chinese researchers, according to the post by Học lập trình cho người mới bắt đầu. The paper argues that humans hand-writing the programs wrapped around AI models is gradually becoming an obsolete skill.

2. How does the AI loop that generates a harness work? A small 4B model reads the errors that occurred and proposes 8 edits; only the highest-scoring edit on an independent test question set is kept, while the large model performing the main task is kept completely unchanged.

3. What is wrong with human-written harnesses, according to the paper? They are only a guess at one moment in time, deployed and then left running for months without being re-read or evaluated for effectiveness.

4. What human role remains? Designing the scoring set — the collection of questions used to grade the harness. Frameworks become continuously replaceable, while designing the test becomes the core job.

5. Has this article verified the paper? No. The article presents the paper's argument as described in the Facebook post and the summary image shared by Mr Buzz; Wotech has not read or independently verified the full text of the study.

This article was rewritten from a post by the page Học lập trình cho người mới bắt đầu on 15 November 2026, summarizing the paper "The End of Harness Engineering". The arguments, figures, and mechanisms follow the original post and the summary image by Mr Buzz; Wotech has not independently verified the full text of the study.