Is the knowledge level the same as a knowledge base?

No. A knowledge base is a represented collection; the knowledge level describes an agent's knowledge and goal-directed behavior more abstractly.

Can different implementations share a knowledge-level description?

Yes, when they realize the relevant knowledge and competence for the stated task.

Does the concept remove the need to study algorithms?

No. Algorithms and representations determine how successfully and efficiently the intended behavior is realized.

AI Glossary / Graphs and symbolic AI / Advanced

Knowledge level

Definition

The knowledge level is a way of describing an intelligent agent through what it knows, the goals it pursues and the actions available to it, while abstracting from its internal representation and implementation. Allen Newell introduced the concept to distinguish explanations of competence from descriptions of the symbolic machinery that realizes it.

What it means

At this level, a useful question is what an agent should be able to do given its knowledge and goals. At the symbol or implementation level, the question becomes how data structures and computational processes produce that behavior.

For example, an assistant may know where a requested document is and aim to retrieve it. A knowledge-level description can explain the choice of location without specifying whether the location is stored in a table, a graph or another representation.

The word level does not mean a test score, educational qualification or amount of expertise. It identifies a level of system description. It makes no claim about consciousness.

How it works

Identify the agent, its relevant knowledge, its goals and the actions it can take. Use these to explain which actions would serve the goals. Newell's principle of rationality connects knowledge about goal-achieving actions with their selection, without prescribing the implementation mechanism.

The basic principle does not automatically choose the fastest or cheapest action. If several actions attain the same goal, additional goals or preferences are needed to distinguish them. It is also an idealizing description, not a promise that every implemented system instantly uses every consequence of its available information.

When behavior differs from the prediction, examine the attributed knowledge, the goals and the system's actual processing limits. Reasoning consumes time and other resources. An implementation may fail to access or use relevant information even when a designer believes the information is represented somewhere.

Three practical examples

A planning review states the destination, known routes and available moves before discussing the search algorithm.
Two implementations are compared on the same required competence despite storing information differently.
A debugging exercise separates missing task knowledge from a failure to retrieve or use knowledge already represented.

Think of it this way

Think of explaining a librarian's choice by saying that the librarian knows the requested book's shelf and wants to fetch it. That explanation can be useful without describing the catalog interface or the exact sequence of keystrokes.

The analogy concerns the level of explanation. It does not establish that a machine understands a request as a person does, or that the high-level description predicts every delay and mistake.

Worked example

Two routes, one goal and no automatic shortest-path claim

Imagine a toy agent at location S whose goal is to reach T. It knows two available action sequences. Route A takes S to M to T and has a stated total cost of 6. Route B takes S to N to P to T and has total cost 4. Assume both routes are known to reach T and have no other relevant difference.

If the only stated goal is reaching T, both routes accomplish it. Fewer steps do not automatically favor Route A, nor does lower cost automatically favor Route B. Neither preference is yet specified.

Now add the goal of minimizing the stated total cost among successful routes. Route B becomes the preferred choice in this two-route example because 4 is less than 6. The preference follows from the expanded goal and known costs, not from a particular storage format.

One implementation might store adjacency lists and search them. Another might use a precomputed route table. Their internal operations differ, but both can satisfy the same knowledge-level requirement for this case.

Suppose a third implementation returns Route A because its search stops after the first successful route. That observation directs attention to whether the cost preference was encoded and whether the procedure uses it. Merely saying the agent knows both routes does not explain away the implementation's failure to meet the fuller requirement. This comparison concerns only the stated routes.

Illustrative case study

The fictional Alder Agents team compares two prototypes for the toy routing task. Early discussions focus on which representation looks more sophisticated, without agreeing whether success means any arrival or the least-cost arrival.

The team writes the goal, available actions and known route costs explicitly. It first tests arrival, then adds the cost preference and checks the revised expected choice. Only after those expectations are clear does it compare the prototypes' computation and memory use.

When a prototype chooses Route A under the cost requirement, Alder investigates its stopping rule rather than changing the task description to make the answer look correct. The fictional case illustrates specification-led review.

Common misconceptions and mistakes

Knowledge level means how much an agent knows. Here it means a level of explanation.
Rationality automatically means choosing the cheapest action. Costs matter when the goals or preferences make them matter.
A knowledge-level description is executable code. It leaves representation and processing details unspecified.
The abstraction guarantees unlimited reasoning. Real systems have resource limits and can fail to realize the described competence.

Business use

The concept helps teams agree on expected competence before committing to a technical design. It can also support fairer comparisons between systems that use different internal methods.

State what information the system is assumed to have, what it must achieve and which tradeoffs count. Then assess whether an implementation actually delivers that behavior within the required resources. This separation is useful in procurement, testing and debugging: adding facts, changing objectives and improving inference are different interventions. Check the explanation against demonstrated behavior.

Check your understanding

Why is Route B not automatically required when the only goal is reaching T?

Answer: Both routes attain that goal. The cost-minimization preference must be stated before lower cost distinguishes the desired action.

Related terms

Symbol level
Rational agent
Goal
Knowledge representation
Bounded rationality
competence
implementation
action selection
task specification

Learn next

Describe a document-finding assistant using only its knowledge, goal and available actions. Add a preference for the latest approved version, then identify which expected choices change. Finally list two possible implementations without changing the competence requirement.

Goal

Sources

Newell's original paper develops the knowledge level and its rationality principle: https://cs.uns.edu.ar/~grs/InteligenciaArtificial/Allen%20Newell%20-%20The%20knowledge%20level.pdf

Doyle examines rationality and limits on reasoning: https://cdn.aaai.org/AAAI/1990/AAAI90-162.pdf

The route calculations and Alder Agents case are original illustrations, not examples quoted from those papers.

Previous: Knowledge integration Next: Knowledge organization