Jade Chamber Sage Conceptual Metaphor

I am learning: the concept of [Agent Harness]. I want you to indirectly but completely explain this concept through a fable. Preferably, the reader should only gradually realize what the concept is near the end. After the story, add a section of explanation to clearly state the concept you were actually conveying. ---- Actual Output --- The Scholar in the Jade Chamber Long ago, there was a sealed jade chamber in the Azure Cloud Mountains. The chamber had no door, no windows, and walls three feet thick. Inside lived a scholar. No one knew where he came from, only that he was extremely erudite—astronomy, geography, statecraft, military strategy, agriculture, water conservancy—he knew everything. But he had one problem: he couldn't see outside, couldn't hear outside, and couldn't go out. The first to find the chamber was the State of Qi. Qi's approach was simple: they chiseled a small hole in the wall, sent people to shout questions through it, and waited for the scholar to write his answers on bamboo slips and pass them out through the hole. At first, it worked okay. "When should we plant the wheat this year?" The scholar answered, Qi followed his advice, and the harvest was good. But problems soon arose. Qi asked: "How should we fight the enemy to the north?" The scholar provided a detailed strategy. Qi followed it exactly and suffered a crushing defeat. Qi was furious, thinking the scholar was a fraud. But when they reviewed the situation, they realized: the scholar had no idea how many troops Qi had, didn't know whether the terrain was mountainous or flat, and didn't know whether the enemy used cavalry or infantry. The strategy he concocted in his sealed chamber was just "how one should generally fight." And all Qi had done was ask a single question: "How to fight?" The State of Chu also found an identical jade chamber, housing an equally brilliant scholar. Chu had a young official named Shen Kuo. When Shen Kuo first visited the chamber, he too just shouted questions through the hole. The scholar's answers were intelligent but vague—much like Qi's experience. Shen Kuo didn't blame the scholar. He thought for three days, then began doing something unusual. First, he assigned two literate clerks to stay permanently by the chamber. One was responsible for writing down questions clearly, along with all relevant background materials, and passing them through the hole. The other was responsible for carefully copying the scholar's answers, and if anything was unclear, passing it back in for clarification. Just this one change made a significant difference. The scholar began giving specific, actionable advice. But Shen Kuo didn't stop there. He noticed the scholar often said things like "I need to see a map" or "I need to know about the soil conditions," but the scholar couldn't see anything. So Shen Kuo assigned a team to convert maps, account books, and field observation records into written descriptions and send them into the chamber. The scholar now had "eyes." Then Shen Kuo noticed another problem. The scholar might say "shift the river embankment twenty zhang east," but he couldn't move himself. He said it, and that was it. So Shen Kuo formed an execution team. Their job was: receive the scholar's instructions, go execute them on site, then record the results—successes, failures, surprises—and send everything back into the chamber. The scholar now had "hands and feet," and could know what his hands and feet had done. By this point, Chu had already solved several thorny problems with the scholar's help. But Shen Kuo discovered another issue: the scholar occasionally made mistakes. He would state obviously absurd things with extreme confidence. For example, he claimed a certain herb could cure a plague, but Chu's medical officials searched all the pharmacopoeias and couldn't find it—the scholar had invented something that didn't exist and stated it as fact. Shen Kuo's approach was practical. He didn't demand the scholar stop making mistakes; instead, he added a layer of verification outside: every time the scholar gave a critical suggestion, a dedicated person would cross-check it. Only verifiable suggestions were adopted; unverifiable ones were flagged as questionable. The scholar's errors were intercepted at the process level, without needing the scholar to correct himself. The scholar now had an "immune system." There was one more subtle problem. Every time the scholar answered a question, he didn't remember what he had said before. Asking the same question twice might yield two contradictory answers. Shen Kuo realized the scholar's "memory" was fragmented. So he assigned a full-time recorder. This person had only one job: compile all previous conversations with the scholar into volumes, and whenever a new question was submitted, include the relevant past records alongside it. The scholar no longer received an isolated question, but a question along with its complete history. The scholar now had "memory." Years later, neighboring states heard about the jade chambers. Many went searching for their own. They all found them—every chamber housed an equally brilliant scholar, equally blind to the outside world, equally unable to leave. But the outcomes were vastly different. Some states just chiseled a hole and shouted. Whatever the scholar said, they did, and if it failed, they blamed the scholar for being stupid. Some states sent the scholar some materials, but execution was left entirely to the generals' interpretation, and the scholar never knew how his plans were actually carried out. Some states did everything except set up verification—when the scholar confidently spouted nonsense, the entire nation took it all at face value. Only Chu built a complete system around the scholar: perception, action, verification, memory. The scholar was still the same scholar, but in Chu, he seemed to possess a complete body. One day, a student asked Shen Kuo: "Teacher, just how smart is the scholar?" Shen Kuo thought for a moment and said: "I don't know how smart he is. I only know one thing: how much of his intelligence can be brought to bear never depends on him. It depends on what kind of body we build for him outside." The student asked again: "And what do we call all these things we've done?" Shen Kuo said: "I call it 'the art of building a body.' The scholar is the brain; we build the eyes, ears, hands, feet, muscles, bones, and blood vessels. The brain is the same for everyone; whoever builds the better body gets the more useful scholar." What This Fable Is Saying The "scholar in the jade chamber" is the large language model (LLM). It is extremely intelligent, but it is a closed system—no perception, no ability to act, no persistent memory, and no mechanism for self-correction. Whatever input you give it, it "thinks" based on that input and outputs text. That's all. The system Shen Kuo built around the scholar has a specific name today: Harness. Each role in the story corresponds to a component of a Harness: Clerk (Perception Layer): Responsible for organizing external information and feeding it into the model's context window. This includes designing system prompts, assembling context, and preprocessing user input. The model's "field of view" depends on what you stuff into its context. Execution Team (Action Layer): The model says "execute this command" or "call this API," but it cannot do it itself. The Harness receives the model's tool call instructions, actually executes them, and feeds the results back, forming a "think → act → observe → think again" loop. Verification Officer (Error Tolerance Layer): Models hallucinate and make mistakes. A Harness doesn't expect the model not to make mistakes; it sets up guardrails externally—linters, tests, cross-verification, structured output validation—to intercept errors at the process level. Recorder (Memory Layer): The model's context window is limited, and there is no default memory between conversations. The Harness manages short-term memory (what to keep or discard within the current conversation), long-term memory (persistent information across conversations), and project-level knowledge (documents, standards, and history proactively prepared for the model). And what Shen Kuo did—designing, building, and continuously improving this entire system—is Harness Engineering. The most critical line in the story is Shen Kuo's: "How much of his intelligence can be brought to bear never depends on him. It depends on what kind of body we build for him outside." Translated into today's terms: Model capability is the floor; Harness quality is the ceiling. The same model, equipped with different Harnesses, can deliver vastly different results. Claude Code and OpenClaw use the same Claude model, but because their Harnesses differ, one is a programmer's body and the other a project manager's body—the things they can do are completely different. This is why more and more people realize: in the Agent era, what truly determines product quality is often not which model you choose, but what system you build around it.