Summary of Key Points
The focus of competition in the AI industry is shifting from “who has the largest model” to “who can best leverage that model”—and the key to doing this is Harness. It’s not just a simple model wrapper; rather, it’s an intelligent system engineering framework that transforms AI from being capable of “chatting” into something that can reliably perform actual tasks. Industry leaders like Jensen Huang believe that the core competitiveness of future companies will lie in Harness, as it turns large models (which are like fast horses) into truly productive tools. Both large companies and open-source communities around the world are vying for control of Harness, and even ordinary individuals can start developing basic versions of it.
What is Harness? — Not Just a Wrapper, but an “Operating System” for AI
You can think of a large model as a fast horse, and Harness is the complete set of equipment needed to control it: the reins to direct its direction, the saddle to ensure stable operation, the brakes to prevent errors, and the dashboard to monitor its performance. It’s different from common Agent Frameworks, which are more like “recruitment manuals” that teach you how to create an AI employee. Harness, on the other hand, acts more like an “enterprise management system” that assigns tasks, provides tools, tracks progress, and oversees execution, enabling the AI employee to work efficiently over the long term.
From a technical perspective, Harness is similar to the Windows operating system on a computer: without it, even the most powerful chip is just a collection of hardware; with it, the AI’s capabilities can be fully utilized. For example, the 510,000 lines of code leaked by Anthropic include six core components of Harness:
- Multi-layer prompts: These set rules for the AI (for instance, “You are a market analyst; you can’t talk nonsense”).
- Tool manuals: These instruct the AI on how to use external tools (such as data retrieval or coding).
- Task execution loops: These ensure tasks are carried out step by step to avoid unnecessary confusion.
- Memory management: This helps the AI remember previous tasks so it doesn’t forget them.
- Sub-agent division of labor: The main AI assigns smaller tasks to specialized sub-agents (e.g., an AI for coding and another for testing).
- Verification mechanisms: These check whether the AI’s results are accurate (e.g., by verifying that code actually works).
Why Does Jensen Huang Value Harness So Much? — Because It Makes AI Profitable
Jensen Huang, the CEO of NVIDIA, focuses on Harness because it transforms large models into productive tools that companies are willing to invest in. Data speaks for itself:
- LangChain experiment: With the same GPT model, adding Harness increased task completion rates from 52.8% to 66.5%.
- Nemotron model: Using Harness reduced evaluation costs by tenfold (from $43 to $4.5).
- Claude model: With Harness, it’s now possible to generate usable software, whereas before only semi-finished products were created.
The more widespread Harness becomes, the more GPUs companies will need to power their AI systems, as each AI “employee” relies on GPUs to function effectively. Huang’s strategy is to integrate Harness into business processes, thereby increasing GPU sales.
The Battle for Harness Has Begun — Big Companies and Open-Source Communities Are All Involved
Companies and open-source communities around the world are competing to dominate Harness, just as they did in the past for control of operating systems:
- OpenAI: In early 2026, they used Harness (a file called AGENTS.md) to have Codex generate 1 million lines of code in five months.
- Anthropic: The leaked code revealed 44 unreleased Harness features, indicating significant investment.
- Chinese companies: DeepSeek has dedicated a team to developing Harness, and Tencent’s WorkBuddy is being tested internally with 2,000 employees before being released to the market.
- Open-source communities: Projects like OpenClaw prevent AI from revealing internal thoughts, and Hermes makes AI more intelligent over time—these are all examples of open-source Harness developments.
Whoever masters Harness will control the future of AI applications.
Even Ordinary People Can Develop Harness — Just 200 Lines of Code Are Enough
You don’t need to write hundreds of thousands of lines of code; the smallest version of Harness requires around 200 lines. For example, to have AI analyze NVIDIA’s news for a weekly newsletter:
- Without Harness: You’d have to ask the model directly, and it might provide random or incomplete information.
- With Harness: You can break down the task into three steps: 1. Find the latest NVIDIA news; 2. Filter out important content; 3. Organize it into an outline. Harness handles the coordination of these steps, while the model focuses on the actual work.
Getting started is simple:
1. Write an AGENTS.md file (200 lines of Markdown) that defines the AI’s role, task rules, and prohibited actions.
2. Use open-source tools (like LangChain) to build the basic components (memory management, task execution).
3. Guide the AI to follow a structured process rather than acting on its own.
Conclusion: The Future Competition Will Focus on Harness
In the future, people might not ask, “What model are you using?” but rather, “How well is your Harness designed?” Large models will be accessible to everyone as basic infrastructure, but a company’s own Harness will be the real differentiator. It determines whether your AI systems can work reliably and generate profits for your business.