The Knowledge Stack — Detail

Four Layers of Agent Knowledge

The Complete Knowledge Architecture

The Gap Skills Fill

Tools vs Procedures

MCP says: "You can call the GitHub API."

Skills say: "When a user asks for a PR, use the GitHub API with these branch conventions, this review process, and these CI checks."

MCP provides capability. Skills provide the judgment of when and how to use it.

Synergy

Better Together

Skills + MCP = tools + procedures
Skills + RAG = procedures informed by current facts
MCP + RAG = grounded tool usage
All four = complete agent

No single layer is sufficient. The stack is the product.

Fine-tuning

Baked Into Weights

Knowledge embedded during training. Expensive to update (weeks/months). Not composable or portable. Cannot be version-controlled separately.

Base Capabilities Language Understanding

Update: weeks | Portability: none

RAG

Retrieval at Query Time

Fetches relevant documents from a vector store or search index at query time. Fast to update (minutes). Medium portability (depends on embedding model).

Documentation Code Search Current Facts

Update: minutes | Portability: medium

MCP

Protocol for Tool Access

Standardized protocol for connecting AI agents to external tools. APIs, databases, file systems. Instant updates. High portability (protocol-based).

API Calls DB Queries File Ops

Update: instant | Portability: high

Skills

Markdown Procedures

Step-by-step instructions in plain markdown. Instant updates (edit and save). Highest portability (works on any platform that reads markdown).

Workflows Conventions Team Practices

Update: instant | Portability: highest

Detailed Comparison

Layer-by-Layer Analysis

Dimension	Fine-tuning	RAG	MCP	Skills
Knowledge type	Weights	Facts	Tools	Procedures
Update speed	Weeks	Minutes	Instant	Instant
Portability	None	Medium	High	Highest
Context cost	Zero	Medium	Low	~100 tok
Composability	No	Limited	Yes	Yes
Version control	No	Partial	Yes	Git-native
Human-readable	No	Varies	Config	Markdown

Hybrid Architectures

Combining Layers for Maximum Capability

Skill-guided RAG

A skill defines what to retrieve and how to use the results. "Search docs for X, extract Y, format as Z."

Skill> RAG Query> Formatted Output

MCP-powered Skills

Skills orchestrate MCP tool calls in a defined sequence with error handling and rollback.

Skill> MCP Tool> MCP Tool> Result

Fine-tuned Skill Matching

Fine-tune the base model to better match user intent to skill metadata. Improves Tier 1 accuracy without loading more context.

Fine-tuned Model> Better Tier 1 Match

Full Stack Agent

All four layers working together. The fine-tuned model understands intent, RAG provides facts, MCP provides tools, Skills provide procedures.

Model+ RAG+ MCP+ Skills= Complete Agent

When to Use What

Decision Guide

Is it a fact?

Use RAG

"What is the current API limit?"

Is it a tool?

Use MCP

"Call the Stripe API"

Is it a procedure?

Use Skills

"Deploy with our CI/CD process"

Is it fundamental?

Fine-tune

"Understand TypeScript generics"

Cognitive Science Mapping

Human Memory Types in AI

Memory Type	Human Example	AI Equivalent	Technology
Semantic	Knowing that Paris is in France	Factual knowledge	RAG
Episodic	Remembering last Tuesday's meeting	Conversation context	Chat history
Procedural	Knowing how to ride a bicycle	Step-by-step workflows	Skills

A surgeon knows anatomy (semantic), remembers a difficult case (episodic), and knows how to make an incision (procedural). Skills are the procedural memory that AI agents were missing.