What is Ontology ? Ontology for Agents

Summary

An ontology is a formal, machine-readable map of the important things in a software system, their properties, and how they relate. Unlike a glossary, which defines terms, or a folder structure, which organizes files, an ontology captures connections—for example, which endpoints a service exposes, which other services it depends on, which tables it owns, and which events it emits.

Ontologies are especially useful for AI agents. An agent exploring a codebase without architectural context may be able to read the code but not understand which parts are essential or how the system fits together. A structured map gives the agent that context and can improve its reasoning.

Ontologies can also help teams without AI agents:

• Assess change impact: Trace what might be affected by modifying a system component.
• Build features faster: Find where existing concepts are implemented and what can be reused.
• Support integrations: Give external systems a clear view of business concepts such as orders, customers, and payments without exposing internal file structures.

Five Steps for Building an Ontology

1. Define core concepts from the top down. List roughly 10–30 important business concepts in the domain. Focus on what the system represents, not just its code artifacts.
2. Choose a simple representation. Start with structured Markdown files stored and versioned in the repository; a graph database or formal ontology standard is not required at the outset.
3. Extract relationships from the codebase. Use dependency graphs, API specifications, and optionally an LLM to draft entities and their relationships. Have people review and correct the draft.
4. Link concepts to actual code. Each ontology entry should point to relevant paths or modules. Without those links, the ontology risks becoming detached documentation rather than a practical tool.
5. Keep it current. Add a check to the pull-request process asking whether a change affects an ontology concept. A stale map can mislead people who rely on it.

The central lesson is that better-mapped codebases provide stronger context for both people and AI agents. A useful ontology can begin with simple Markdown files and grow alongside the code.

Speakers and Sources

• Presenter/narrator: One speaker throughout the video; the channel metadata identifies the channel as Azim Shaik.