Skip to content

Architecture Map — Every Keyword's Place and Role in One View

LLM, Agent, Tool Calling, MCP, Skills, Workflow, Memory, Knowledge Graph, GraphRAG, and RAG are not parallel concepts of equal rank. Each has its own post.

About This Document

In enterprise AI adoption, a crowd of keywords tends to be presented as "options" at the same level. In reality, they are parts with different roles, arranged in layers with dependencies. This page provides that overall view, and doubles as a signpost to the section of this site where each keyword is covered.

Audience: First-time readers of this site; developers and architects who want the big picture of AI adoption

1. How the Terms Relate

TermIn One PhraseContrasting Concept
LLMA function that predictsAgent (a subject that loops)
AgentAutonomous judgment loopWorkflow (fixed procedure)
Tool CallingThe calling mechanismMCP (its standard protocol)
MemoryWhat to rememberContext Window (volatile)
Knowledge GraphStructured relationsVector DB (similarity)
GraphRAGRetrieval that walks relationsPlain RAG (fragment retrieval)

2. Resource Types × Access Paths

Most of these keywords can be organized as pairs of "which resource, accessed by which means". The Agent is not a peer of the others — it sits above them as the orchestrating layer.

Resource TypeNatureAccess PathNotes
Document knowledgeUnstructured, staticRAGRead-only "search by meaning". Extends to GraphRAG for cross-document questions
Business dataStructured, dynamicDB (SQL / Semantic Layer)Read access for "exact values". The LLM only generates the query
Business operationsSide effectsAPIWrite and execute — includes irreversible operations, so permission design is mandatory
Relational knowledgeGraph-structuredKnowledge Graph / Memory"Who owns what, what depends on what". Retrieved via GraphRAG
Multi-step executionCombination of the aboveAgent / WorkflowThe orchestration layer spanning resources. Fixed steps → Workflow; judgment needed → Agent

IMPORTANT

The value of this classification is that "read vs write" separates naturally. RAG and DB are reads (safe, idempotent); API is operations (side effects, permissions required). When you grant an Agent permissions, this boundary becomes the risk boundary. See Permission vs Authority.

NOTE

MCP does not get its own row in this table — it runs across it, as the connection standard unifying RAG / DB / API access. Skills are the static knowledge and procedures the Agent layer consults; they belong to "defining the Agent's behavior", not to access paths.

3. Data Flow — a Cycle, Not a One-Way Street

The two return flows are the point. The Business Process generates new Data, and the Agent's execution experience accumulates in Memory, growing the Knowledge. Without these return flows, you fall back into the scatter-gather problem of "re-researching everything from scratch every time".

WARNING

If your data is fundamentally scattered, AI is not the solution. Put RAG or an Agent on top of scattered, dirty data and it will only reproduce the scatter faster. Data curation comes first — unifying relations in a Knowledge Graph, unifying metric definitions in a Semantic Layer.

4. Keyword → Section of This Site

KeywordCovered InRole
LLM (structural constraints)Sister site: understanding-llmThe "Why" bookshelf
Agent / Sub-agent / A2AAgentsTaxonomy and design of executors
Tool Calling / MCPMCPConnection as an implementation mechanism
SkillsSkillsStatic knowledge and procedures as an implementation mechanism
WorkflowWorkflowsPatterns for fixed procedures
Memory / Knowledge GraphConcepts 08Concepts of memory and knowledge integration
RAG / GraphRAGThis section (page in preparation)Access design for document knowledge
Semantic LayerMCP / Semantic LayerDesign discipline for structured data access
Doctrine (decision criteria)Concepts 07Constraints, purpose, decision criteria
Permission / AuthorityStrategySeparating permission from authority

TIP

When unsure which means to pick, three axes decide it: freshness (static → RAG, dynamic → DB/API), amount of judgment (none → Workflow, much → Agent), and state of the data (dirty → curate first; AI comes last).

🔗 Going Deeper: Why LLMs Need an External Information Foundation

This page covered the structure (What/How) of information architecture. To understand why an LLM alone is not enough — from the LLM's structural constraints — see the sister site.

  • understanding-llm (top page) — The eight structural constraints (Context Rot, Knowledge Boundary, etc.) that make external references necessary

Previous: Overview (Information)

Last updated: August 2026

Released under the MIT License.