Why Knowledge Bases Alone Can’t Empower AI Agents: The Need for Actionable Experience

Large models possess extensive factual knowledge, but when deployed in real‑world scenarios they often fail to complete specific tasks. The core contradiction in the era of self‑evolving AI has shifted from "having knowledge" to "being able to turn knowledge into action" (the article’s opening observation).

1. Skills are compressed action experience, not static manuals. A skill comprises procedural knowledge, contextual judgment, pattern recognition, automated execution, and reflective adjustment. It is high‑level implicit knowledge that cannot be fully expressed as a list of facts; expert intuition is the rapid recognition derived from massive experience.

2. Skills have no single standard answer; they must respect each model’s cognitive endowment. Different base models (e.g., Claude 4.6 vs. Qwen 3.5) encode the same task differently when using GenericAgent, producing either a transferable method template or a concrete operation script. The choice of representation depends on the model’s reasoning habits, tool understanding, and long‑term planning ability.

3. Agents become new carriers of human skills. Under expert supervision, agents explore environments, expose failures, receive corrections and demonstrations, and iteratively distill these experiences into reusable skills. This transforms expert knowledge from a personal, tacit asset into a machine‑learned capability.

4. Knowledge remains essential, but its content outweighs its form. Traditional knowledge graphs, database schemas, and rule bases still provide consistency and verifiable reasoning for strict domains. However, agents also need natural‑language instructions, tables, code snippets, and operational procedures that can be directly understood and executed.

5. Knowledge that lives is produced in action. The article illustrates a Baidu map‑measurement task: the model knew to zoom the map but missed the crucial step of positioning the mouse over the target, causing failure. Repeated practice captured this micro‑action, turning an accidental omission into a stable skill.

6. Practice is the production site of knowledge, not merely its application. Each execution must record four signals: the action taken, environmental change, success/failure outcome, and diagnostic reasoning. Only with this process evidence can agents compare strategies, identify stable causal factors, and form reusable procedures.

7. Knowledge construction should target blind spots rather than bulk. Private‑domain gaps—such as internal error‑code meanings, service topology, user preferences, and historical fault handling—are far more decisive for task success than the sheer volume of public knowledge. Building an "executable experience library" that records applicability, conditions, risk level, and validation evidence is essential.

In summary, self‑evolving AI transforms knowledge engineering from a one‑time description of the world into a continuous cycle of skill creation, validation, and selective forgetting. Success is measured by reduced error rates, the ability to acquire new skills under supervision, and the capacity to adapt or discard outdated experience.