AI Agent Engineering Handbook
"The best AI engineers understand both the models and the engineering."
This knowledge base builds a complete technical loop from LLM fundamentals to production AI agent systems.
The Core Formula
Agent = Model (Brain) + Prompt (Instruction) + Memory (RAG/Context) + Tools (MCP) + Planning (Architecture)
1. System Architecture Overview
This diagram shows how the 7 modules logically depend on each other:
How the modules connect:
- LLM Foundational provides computing and reasoning foundations
- Prompt & Context are the media for interacting with models
- RAG provides static knowledge support for models
- MCP provides dynamic tool support for models
- Agents orchestrate and coordinate all above components
- Ops & Security runs through the entire lifecycle
2. Module Synopsis
| ID | Module | One-Liner Definition | Key Technologies & Keywords |
|---|---|---|---|
| 01 | LLM Foundational | Understanding the "brain" mechanism, training pipeline, and physical limitations | Transformer, Attention, Pre-training, RLHF, Tokenization, Inference Params (Temp/Top-P) |
| 02 | Prompt Engineering | Writing "instruction code" to elicit reasoning and standardize output format | Chain-of-Thought (CoT), Few-shot, ReAct, XML/JSON Output, Persona |
| 03 | RAG | Augmenting models with external "library" to solve hallucinations and inject private data | Vector DB, Embeddings, Chunking, Hybrid Search, Grounding, Self-Querying |
| 04 | Agents | Evolving from "chat" to "action" with planning, reflection, and tool use | Orchestration, Loop Control, Reflection, Router, Multi-Agent (Supervisor/Hierarchical) |
| 05 | MCP | Model Context Protocol - standardized AI connection (USB-C) decoupling models from tools | Host/Client/Server, Resources, Tools, Prompts, JSON-RPC, Stdio/SSE |
| 06 | Context Engineering | Managing model "attention" window and long/short-term memory to prevent overload | KV Cache, Context Window, Short/Long-term Memory, Information Compression |
| 07 | AgentOps & Security | Converting demos to production applications with safety, observability, and evaluation | Eval (LLM-as-a-Judge), Prompt Injection, Docker Deployment, Tracing |
3. Learning Paths
Choose your path based on your development goals.
The Builder Path (Practical Developer)
Goal: Quickly build a Java AI Agent that can access the web and query databases.
Recommended Order:
- 05 MCP: First understand how to write a tool (Server)
- 04 Agents: Learn how to make the model call this tool
- 02 Prompt: Optimize instructions for more accurate calls
- 07 Ops: Deploy to Docker (refer to Brave Search case)
Focus: Rapid iteration, working code, production deployment
The Architect Path (Architect/Researcher)
Goal: Design complex enterprise multi-agent systems.
Recommended Order:
- 01 Foundational: Understand model capability boundaries
- 04 Agents: Design multi-agent collaboration patterns
- 06 Context: Design memory systems to support long workflows
- 03 RAG: Plan enterprise knowledge base integration
Focus: System design, scalability patterns, architectural trade-offs
4. Quick References
Essential resources for common tasks - avoid deep-diving into documentation.
Standard Agent System Prompt Template
MCP Server Standard Code Structure (Java/Spring)
RAG Chunking Strategy Cheat Sheet
Recommended LLM Parameters
| Parameter | Conservative | Creative | Coding |
|---|---|---|---|
| Temperature | 0.0 - 0.3 | 0.7 - 1.0 | 0.1 - 0.2 |
| Top-P | 0.9 | 0.95 | 0.9 |
| Max Tokens | 1024 | 2048 | 4096 |
| Frequency Penalty | 0.0 | 0.3 | 0.0 |
5. Navigation Guide
Core Modules
- LLM Foundational - Transformer architecture, training, inference, limitations
- Prompt Engineering - CoT, few-shot, ReAct patterns, output formatting
- RAG - Vector databases, embeddings, retrieval strategies, grounding
- Agents - Orchestration, multi-agent systems, planning, reflection
- MCP - Protocol specification, server implementation, tools, resources
- Context Engineering - Context windows, memory systems, optimization
- AgentOps & Security - Deployment, monitoring, safety, incident response
Additional Resources
- Java & AI Internship Guide - Career development, practical skills
6. Key Concepts at a Glance
Token Economics
- 1 token ~= 0.75 words (English) ~= 4 characters
- Context window = maximum tokens per request (varies by model)
- KV cache = cached previous tokens for faster generation
The RAG Pipeline
Query -> Embedding -> Vector Search -> Context Assembly -> LLM -> Response
Agent Decision Loop
Observe -> Reason -> Act -> Observe -> Reason -> Act ...
MCP Connection Model
Host (App) <-> Client (Protocol) <-> Server (Tool/Data)
7. Common Patterns
Pattern 1: ReAct Agent
Thought: [Analyze the situation]
Action: [Call a tool]
Observation: [Review result]
Thought: [Plan next step]
Action: [Continue or finish]
Pattern 2: Router Agent
Classify Query -> Route to Specialist Agent -> Aggregate Results
Pattern 3: Hierarchical Agents
Supervisor Agent -> Worker Agents -> Report Back -> Synthesize
8. Production Checklist
Before deploying to production:
- All tools have proper error handling
- Sensitive operations require human approval
- Comprehensive audit logging enabled
- Kill switches implemented and tested
- Rate limiting configured
- Cost controls in place
- Monitoring dashboards active
- Incident response procedures documented
- Security review completed
- Load testing performed
Get Started
New to AI engineering? Start with LLM Foundational to understand how models work, then move to Prompt Engineering to learn effective prompting patterns.