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". In this diagram, Agent / Workflow sit on the orchestration side. In the II.1 five layers, however, Agent is one ownership post — not the top of a resource stack.

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. Relation to the Five Layers ​

The diagrams on this page are a map of resources, access paths, and circulation. The main map of this book is the II.1 five layers. The axes differ. Do not read this page as another name for the five layers.

Five layersHow it appears on this page
DoctrineNot a node in the diagrams. The measure of purpose, prohibitions, and priority. See the table below and III.3 Doctrine
AgentOn the orchestration side. In the five layers it owns understanding and assignment — it is not an upper stack that owns the other layers
SkillsNot an access path. Stable knowledge and procedures the Agent reads. Kept off the diagram
MemoryRelational knowledge, and one end of the feedback from runs
MCPThe connection standard that cuts across RAG / DB / API

The execution boundary (Harness / Hooks) is none of the five layers. It is a machine interrupt at a point in the run — stop, record, after-steps. It does not add a layer.

NOTE

When unsure, decide who owns it with the five layers first. Then use this page for which resource, read or write. Reverse the order and ownership mixes with connection.

5. Keyword → Section of This Site ​

KeywordCovered InRole
LLM (structural constraints)I.1 Constraint summary / Sister site: understanding-llmWhy and premises
Ownership split (five layers)II.1 Five layers / II.2 PlacementMain map of this book
Doctrine (decision criteria)III.3 DoctrinePurpose, prohibitions, priority
SkillsIII.1 SkillsStatic knowledge and procedures
MCP / Tool CallingIII.2 MCPConnection as an implementation mechanism
Memory / Knowledge GraphIII.4 MemoryHow memory and relations persist
Agent / Sub-agent / A2AIII.5 AgentTaxonomy and design of executors
RAG / GraphRAGThis page §2 (no dedicated page)Read access to document knowledge. For types, see IV.1 Patterns
Semantic LayerMCP / Semantic LayerDesign discipline for structured data access
WorkflowWorkflowsPatterns for fixed procedures
Patterns / limitsIV.1 Patterns / IV.2 LimitsChoosing a type and how far it reaches
Permission / AuthorityPermission vs. AuthoritySeparating permission from authority
HooksHooksHarness-side execution boundary. Not a layer

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

Last updated: August 2026

Released under the MIT License.