Four Levels of AI: Programming, Agents, Judgment, and Panorama

By Scott Shields – Contributing Writer – Capitol Times Media - From Conversations and Material of Zhu Weisha. Learn more about Zhu Weisha here at Capitol Times Media's July Magazine Issue

How AI Handles Certainty and Uncertainty

1. Many AI debates are talking about different things from the start

Asking AI to write a well-defined program is not the same kind of problem as asking AI whether

there is an AI bubble.

Having an Agent complete a task under established rules is not the same kind of problem as asking

who will ultimately benefit from AI.

A major reason many AI debates talk past one another is that these different levels are not separated first.

I think the problems AI encounters in the real world can be divided, broadly, into four levels:

The first level is programming.

The problem is clear, and we already know how it should be solved. The inputs, rules, outputs, and tests of correctness are relatively explicit. AI turns human intent into code, and the resulting program executes those rules rigidly.

The second level is the Agent.

The external environment changes, but we can already specify what should be done under different conditions. An Agent can observe the environment, call tools, and execute tasks, but its actions should remain within pre-established boundaries.

The third level is judgment.

The problem is clear, but the answer is not. “Is there an AI bubble now?” is this kind of question. We can gather evidence, look for counterevidence, and compare explanations, but no program can prove today exactly how the future will unfold. These are problems of uncertain judgment: we may reach probabilistic or conditional conclusions, but we cannot treat them as established facts.

The fourth level is panorama.

This level is harder. We may not even know what the real question is. Take “Who will benefit from AI?” At first, it sounds like a question about which companies will make money.

But once the inquiry opens up, it also involves workers, consumers, governments, countries, capital owners, and even human cognitive capacity and autonomy. In other words, at the beginning we have not even decided what counts as “benefit.” The first two levels have already formed relatively determinate structures. The third and fourth remain structurally uncertain. A major problem in current AI discussions is that these four levels are often mixed together. It is especially dangerous to hand an unresolved third-level judgment directly to a second-level

Agent for execution, or to give an unresolved fourth-level problem to AI and demand an answer. That is the real subject of this essay: when facing different degrees of certainty and uncertainty, what should AI actually do?

2. The real dividing line is how far the problem has become determinate

The four levels differ in one basic respect: how far has the problem itself become determinate? These four levels are not labels for products, professions, or industries. They describe the current state of a problem. The same piece of work may contain problems from several levels at once, and it may move between levels as understanding grows or rules fail.

The first level is the clearest. We already know how to do the task; what remains is accurate repetition. At the second level, the outside world may change, but we already know what action to take when a given condition appears. The third level is different. We know what question we are asking, but we do not yet know what conclusion we should reach. The fourth level goes one step further: we do not yet know what we should really be asking. The four levels can therefore be stated in four short questions: First level: How do we do it? Second level: When do we act? Third level: How should we judge? Fourth level: What is the real question? Uncertainty increases as we move upward.

3. Level One: Program Structure — we already know how

Program structure handles the most determinate class of problems. For example: “Merge two tables by security code; keep the most recent record when there are duplicates; throw an error if fields conflict.” The task may be tedious, but it is already clear. AI can write the program, and once the program exists, the machine can execute it consistently. There is an important point here: A program does not create certainty. The certainty comes from the problem having been specified before the code is written. Now change the request to: “Build me a good stock-analysis system.” Everything changes.

What does “good” mean? What should it analyze? Which data should it use? When should it alert? What counts as an anomaly? Which judgments may be automated? None of this has been settled. If AI starts coding immediately, it has no choice but to guess.

The program may run, but it may not be running the system the user actually intended. In practice, the pattern is consistent: the clearer the intent, the better the AI-generated program; the broader the request, the more gaps appear. AI has not removed the uncertainty. It has silently filled in the missing rules. So the central point of the first level is simple: AI programming can turn rules that are already determinate into determinate execution.

4. Level Two: Action Structure — the world is uncertain, but the action rules are determinate

Agents operate in a more complicated environment than ordinary programs. The outside world does not present the same fixed inputs repeatedly. A self-driving car, for example, does not know whether the pedestrian ahead will suddenly cross the road. But we can specify what conditions require slowing down, what conditions require stopping, where authority ends, and when control must be handed back to a person. That is an action structure. Agents are therefore suited not to every uncertain problem, but to problems in which the environment is uncertain while the rules of action have already become determinate. By “Agent” here, I am not defining a particular technical architecture. I mean a system that has been granted some autonomous authority to observe an environment, call tools, and take action within an established goal, permission set, and action boundary. If an Agent is going to operate in a setting with real consequences, I think it should at least meet several conditions: It knows what the objective is; its capability has been tested; it knows when it must not continue; its permissions are constrained; it can stop when something fails; and its actions can be replayed. Replayability must mean more than the system keeping its own logs. An independent third party should be able to reconstruct what the Agent saw, what rules it relied on, why it acted, and whether it exceeded its authority. Only then can we distinguish between an execution error, bad data, and a flawed action structure. The real question at the second level is therefore not, “Can the Agent do it?” but: “What standard must the Agent meet before it is entitled to do it?”

5. Level Three: Judgment Structure — the question is clear, but the answer is not

The third level is the one most easily confused in current AI use. Consider the question: “Is there an AI bubble now?” We know that investment, demand, utilization, financing, prices, revenue, and cash flow matter. But the future has not happened. We cannot run a program today and receive a final, certain answer. What we need here is not an Agent, but a judgment structure. A sound judgment should answer at least these questions: What exactly are we judging? Which facts support the current view? Which facts cut against it? Can the same facts be explained differently? What are the evidentiary limits? Are important variables missing? Given the evidence now available, what judgment is most reasonable? What new fact would require us to revise it? The third level can therefore be institutionalized. But what is institutionalized is not the answer. What can be institutionalized is the way we form a judgment. The method may be stable; the conclusion cannot be fixed. That is the fundamental difference between the third and second levels. The second level can already say: “If this happens, take this action.” The third level can only say: “Given the evidence available now, this is our current judgment.” That is also why simply adding more Agents does not solve a third-level problem. One Agent can be bullish, another bearish, and several more can vote. That still does not make the future determinate. Five Agents can all be wrong. Different views may compete at the judgment stage, but unresolved disagreement must not be granted action authority. The third level therefore needs a judgment structure, not more Agents.

6. Level Four: Panorama Structure — the real question has not yet been found

The fourth level is the hardest. It often begins not with a clear question, but with a feeling: “Something here does not seem right.” Or: “The prevailing explanation seems too simple.” At this point we have only a vague concept.

We do not yet know the real question, the most important variables, or even the right scale on which to look. Suppose we ask: “Who will benefit from AI?” Used in the ordinary way, AI can produce a list within minutes: chip companies, cloud providers, software companies, robotics firms. It looks complete. But a serious inquiry quickly shows that the question is not only about companies. The beneficiaries may include firms, capital owners, workers, consumers, governments, and countries. And “benefit” may mean profit, wages, consumer surplus, tax revenue, national competitiveness, individual capability, or even cognitive autonomy. What does that tell us? That the original question had not yet fully formed.

The first task at the fourth level is therefore not to define the question immediately, and certainly not to answer it. It is to keep the problem open long enough for different views, facts, counterexamples, and scales to enter. Ordinary search works like this: I know the question, so I search for an answer. Panorama works differently. With panorama, I only know that there may be an important problem. I first open up as much of the possible world as I can, and only then ask where the real question lies.

Our research process usually looks like this: Sweep broadly → find gaps → build a structure → attack the structure → form a judgment → compress it into writing. The “attack the structure” step matters. Once people and AI have built an elegant explanation, they naturally begin to like it. New material then easily becomes evidence for what they already believe. So we deliberately change keywords, look for opposing views, seek out facts that are hardest to explain, and sometimes ask again: were we asking the wrong question from the beginning? Panorama is not about collecting more material. Its real purpose is to stop the first plausible framework from closing the problem too early.

That leads to an important rule for using AI at the fourth level: use AI first to enlarge the world, not to define the world. When should panorama give way to judgment? A practical test is this: the object, scale, and main variables have become relatively stable, and new information mainly changes the conclusion rather than repeatedly changing what we think the question itself is. At that point, the problem has genuinely moved from the fourth level to the third.

7. Why panorama is difficult to scale without AI

People with unusually broad vision have always existed. But such vision usually depended on years of accumulation, cross-disciplinary training, and individual talent. It was difficult to reproduce systematically. Human reading speed, memory, search range, and ability to compare distant fields are all limited. Faced with a genuine fourth-level problem, a person can easily seize the first familiar explanation and then dig ever deeper into it—ending with an excellent study of the elephant’s leg or trunk, but not the elephant.

LLMs and search change more than speed. They sharply reduce the cost of keeping a problem open. AI can rapidly expand dozens of viewpoints, search across disciplines, compare theories side by side, change search terms, find counterexamples, and attack a structure that has only just been formed. That used to be extremely expensive.

AI’s new research value is therefore not merely answering a question faster, but helping us move faster from a vague sense that something matters to a question that is actually worth answering. In this sense, LLMs and search have sharply lowered the cost for ordinary people to expand their cognitive field systematically. But they do not guarantee a genuinely broader view. Without deliberate reframing and counterexample search, AI can just as easily lead us into the first elegant answer.

8. Problem Downshifting: from panorama to program

Once the four levels are placed together, a larger process becomes visible: Panorama Structure → Judgment Structure → Action Structure → Program Structure I call this process problem downshifting. At the fourth level, panorama helps us find the real question. Once the question becomes relatively stable, it enters the third level and we begin to form a judgment. Parts of that judgment that become stable enough to express as rules can then settle into a second-level action structure. As action is further standardized, some of it eventually becomes a first-level program structure.

This process has always existed. In the past it might have looked like this: experts work for years → experience accumulates → rules emerge → processes are standardized → software is finally written. AI is now accelerating this process. It can look more like: vague problem → rapid AI expansion → a human notices a missing piece → AI searches again → judgment forms → the stable part becomes an action rule → AI writes the code → the program executes. The deeper change is this: AI is accelerating the parts that can actually be articulated and verified as they move from uncertain judgment into determinate production processes.

9. But downshifting is not always better

There is an obvious temptation here. If problems can be downshifted, should everything eventually become an Agent and then a program? No. The right principle is: a problem should only be downshifted as far as the evidence allows. Many third-level problems may remain judgment problems for a long time. Many fourth-level problems may need to remain open. Only the parts that have genuinely become determinate should move downward. Otherwise, we get something dangerous: false downshifting.

Suppose the quality of a company’s management is a complex question. To automate management, the organization compresses it into a single KPI. The program then optimizes that

KPI with great precision. But what if the KPI does not represent the company’s real objective? The more accurate the machine becomes, the more stable the error becomes. This is especially dangerous in the LLM era. AI is extraordinarily good at filling in what is missing and making the result look complete.

That creates a new risk that was far less visible in earlier computing eras: In the early computer era, people worried that machines could not solve unstructured problems. In the generative-AI era, we must also worry that machines will make problems that are still unstructured look as though they have already been solved. In the past, when the machine could not do something, the failure was visible. Today AI will answer almost anything. The real danger is that it can look so convincingly complete.

10. A system that can downshift must also be able to upshift

Reality does not obey yesterday’s rules forever. Markets change. Technology changes. Participants learn. Regulation changes. Indicators that once worked can lose their meaning.

Stock markets make this especially obvious. Once a pattern becomes widely known, participants trade ahead of it, and the pattern itself can disappear. A reliable system therefore cannot move only from Level Four to Level Three to Level Two to Level One.

It must also be able to move back upward. If a program or Agent repeatedly behaves abnormally, the problem should return to the third level: was the original judgment wrong? If the measuring stick itself turns out to be wrong, the problem must reopen at the fourth level: were we asking the wrong question?

That is upshifting. A mature AI system should not only know when to automate. It must also know when it has lost the right to continue automating.

11. Form the judgment before issuing the instruction

The four levels differ, but they share one principle. Except for very simple and highly determinate first-level tasks, we should not skip the step of discussion and judgment formation before issuing instructions to AI. At the first level, the main question is whether the specification is clear. At the second level, the main questions are whether the action rules, permissions, and stop conditions are settled. At the third level, the main work is to determine how far the evidence actually supports a judgment.

At the fourth level, the work comes even earlier: find the problem first.

The general pattern of AI use therefore should not simply be: Human writes a prompt → AI executes. A better pattern is: vague intent → human-AI discussion → judgment forms → instruction becomes explicit → AI, Agent, or program executes → result is verified → discussion reopens when necessary. Many failures in AI use are not caused by bad prompts. They happen because execution begins before judgment has been formed. That is dangerous because AI will automatically fill in the missing “judgment,” and we may not know whether those additions are valid.

Conclusion: AI is not one capability, but four different ways of working

People often ask: what can AI do? The question may be too coarse. AI should occupy different positions for different kinds of problems. If we already know how to do something, use a program structure.

If we already know what action should be taken under different conditions, use an Agent. If the problem is clear but the answer remains uncertain, keep it at the judgment level. If the real problem has not yet been found, begin with panorama.

The important thing is not to hand more and more work to AI. It is to ask first: what level has this problem actually reached?

This, I increasingly think, is one of the most basic principles for handling certainty and uncertainty in the age of AI: Not every problem needs an answer. Some problems first need a judgment. And some problems first require us to find the real question.