AI Agent / LLM application engineer

This map threads together everything in the repository. Cross-drill four lenses: the four engineering layers, the interview categories, the classic translations, and the glossary.

Click each node's ▶ to expand the next level. Any blue link jumps straight to the matching HTML chapter, question, or glossary entry. Glossary links also open a popover without leaving the current page.

The reading entrance is the English portal. index.html is the Chinese portal. This page is the review map, not a substitute for the portal.

1. The four engineering layers (the core skeleton)

prompt → context → agent → harness is a four-layer stack of abstractions. Each layer is its own discipline, and a production system needs all four.

🎯 Layer 1 · Prompt engineering Prompt

Core question: How do you get an LLM to produce controllable, high-quality output in a single call?

Transformer foundations — Attention / Multi-head / Position encoding

L1 §1.1 Transformer and attention
Matching questions: A1 LayerNorm vs BatchNorm A2 Multi-head attention A3 RoPE
Core terms: LayerNorm RoPE YaRN GQA/MQA MoE multi-head attention

Training stack — Pretrain → Post-train → SFT → RLHF/DPO/GRPO/RLVR

L1 §1.2 LLM training stack
Matching questions: Category B, 10 questions → all
Core terms: SFT LoRA QLoRA RLHF DPO PPO GRPO RLVR RM catastrophic forgetting

Inference and decoding — Greedy / Top-k / Top-p / Temperature / Beam

L1 §1.3 Inference and decoding
Core terms: temperature logit autoregressive

Classic prompting techniques — CoT / Self-Consistency / ToT / ReAct / Self-Refine / Reflexion / Long CoT

L1 §1.4 Classic techniques
Matching questions: Category C, 9 questions → all
Core terms: CoT Self-Consistency ToT ReAct Self-Refine Reflexion few-shot

Automatic optimization — DSPy / OPRO / APE / prompt versioning

Evaluation and red teaming — LLM-as-judge / Prompt Injection / Jailbreak

L1 §1.6 Evaluation and red teaming
Matching: C7 Prompt Injection I Safety, 6 questions
Core terms: LLM-as-judge Eval prompt injection jailbreak guardrails

🧱 Layer 2 · Context engineering Context

Core question: How do you fit the most relevant information into a limited context window?

Physics of long context — KV cache / Attention sink / Lost in the Middle / RULER

L2 §2.2 Physics of long context · §2.3 RULER
Matching: A5 Lost in the Middle A10 Does 128k actually work? D8 Long context vs RAG
Core terms: context window KV cache attention sink Lost in the Middle

Inference acceleration — PagedAttention / FlashAttention / vLLM / continuous batching / prefix caching

Full-stack RAG — Chunking / Embedding / Hybrid Search / Reranker / Vector DB / RAGAS

L2 §2.4 Full-stack RAG selection
Matching: Category D, 10 questions → all
Core terms: RAG chunking embedding embed models BM25 hybrid search reranker vector DB RAG variants RAGAS

Memory systems — MemGPT / Letta / Mem0 / Zep / working vs episodic vs semantic

L2 §2.5 Memory systems
Matching: Category F, 8 questions → all
Core terms: sub-agent progress file Memory Corruption

Context management — sliding window / summarization / sub-agent isolation / compaction

MCP / A2A protocols — Resources / Tools / Prompts / cross-vendor

L2 §2.7 MCP / A2A
Matching: E5 MCP protocol
Core terms: MCP

🤖 Layer 3 · Agent engineering Agent

Core question: How do you let an LLM decide on its own, call tools, and finish a task across many steps?

Anthropic's 6 patterns — Augmented LLM / Prompt Chaining / Routing / Parallelization / Orchestrator-Workers / Evaluator-Optimizer

L3 §3 Agent architecture
Matching: E1 Workflow vs Agent E2 Six patterns
Required reading: Building Effective Agents (Chinese notes) · original

How agent paradigms evolved — ReAct → Plan-and-Execute → Reflexion → long-horizon autonomy

Agent frameworks — LangChain / LangGraph / LlamaIndex / Haystack / Anthropic SDK

L3 §5 Framework comparison
Core terms: LangChain / LangGraph

Multi-agent collaboration — Supervisor (recommended) / Hierarchical / Swarm (anti-pattern)

L3 §6 Multi-agent
Matching: E4 Choosing a collaboration pattern J5 Conflict handling
Required reading: Multi-Agent Research System (Chinese notes) · original

Tool Use / Function Calling — schema design / granularity / Deny-first / Error as prompt

E8 Tool use design
Core terms: tool use deny-first Computer Use

Four major agent failure modes + Loop of Death — Anthropic's empirical summary

Matching: E6 Four failure modes E10 Loop of Death
Required reading: Effective Harnesses (Chinese notes) · original
Core terms: Loop of Death

Agent evaluation — general benchmarks vs business domains

Matching: Category K, 5 questions → all
Focus: K1 General benchmark landscape K5 Business-domain evaluation

🛡️ Layer 4 · Harness engineering Harness

Core question: How do you engineer the environment so the agent is physically unable to repeat a mistake?
"anytime you find an agent makes a mistake, take the time to engineer a solution such that the agent never makes that mistake again." — Hashimoto

Naming the harness, and where it sits — Hashimoto 2026-02 · boundary with the other three layers

Two core laws + four pillars — strong constraints / self-healing loop · context / specialization / persistence / structure

L4 §3.3 Methodology + supporting infrastructure
Matching: E12 Harness components J4 Stop repeating the same mistake
Core terms: harness AGENTS.md progress file sub-agent deny-first lint type check CI/CD

Supporting infrastructure — MCP / Sandbox / Guardrails

Matching: G7 Sandbox selection I4 NeMo Guardrails I5 Deny-first defense in depth
Core terms: MCP sandbox guardrails

Production-stack reference architecture + maturity model L0–L4 — full 8-layer stack / role matrix / self-diagnosis

L4 §3.4 Production stack
L4 §3.4.6 Maturity model
Required reading: OpenAI Harness Engineering (Chinese notes) · original

2026 model comparison + cost optimization — 8 ROI levers

Scenario selection checklist — internal tools / customer support / coding / multi-agent business systems

2. 92 interview questions · 11 categories

Category	Core coverage	Difficulty	Entry
A. LLM fundamentals (11)	Transformer / RoPE / KV cache / long context / inference optimization / prefill·decode	1⭐ 8⭐⭐ 2⭐⭐⭐	→ A
B. Training / fine-tuning (10)	SFT / DPO / GRPO / LoRA / R1 emergence	1⭐ 7⭐⭐ 2⭐⭐⭐	→ B
C. Prompt engineering (9)	CoT / Self-Consistency / ReAct / automatic prompt optimization	1⭐ 7⭐⭐ 1⭐⭐⭐	→ C
D. RAG (10)	Hybrid / Reranker / GraphRAG / Self-RAG / RAGAS	1⭐ 7⭐⭐ 2⭐⭐⭐	→ D
E. Agent (12) core	6 patterns / multi-agent / failure modes / harness	1⭐ 6⭐⭐ 5⭐⭐⭐	→ E
F. Memory (8)	MemGPT / Letta / Mem0 / progress files / corruption	1⭐ 5⭐⭐ 2⭐⭐⭐	→ F
G. System design (8)	customer support / coding / search / gateway / high concurrency / cost / sandbox / closed loop	0⭐ 1⭐⭐ 7⭐⭐⭐	→ G
H. Evaluation (6)	judge bias / pairwise / benchmark hacking / drift	0⭐ 2⭐⭐ 4⭐⭐⭐	→ H
I. Safety / guardrails (6)	OWASP / Prompt Injection / NeMo / EU AI Act	0⭐ 1⭐⭐ 5⭐⭐⭐	→ I
J. Behavioral / practice (7)	Hashimoto closed loop / 0-to-1 rollout path / debug hallucination	0⭐ 2⭐⭐ 5⭐⭐⭐	→ J
K. Agent evaluation (5)	general benchmarks / trajectory / tool-use / production monitoring / business domains	0⭐ 2⭐⭐ 3⭐⭐⭐	→ K

JD research: 30+ roles in China and abroad → JD Requirements chapter

3. Chinese translations of classic essays

Essay	Core idea		Entry
Building Effective Agents	Workflow vs Agent / 6 patterns / simplicity first	Anthropic 2024	→ Chinese notes · original
Effective Harnesses	4 failure modes and countermeasures / progress files / 5-step startup protocol	Anthropic 2025	→ Chinese notes · original
Multi-Agent Research System	Orchestrator-Workers / sub-agent isolation / Verifier	Anthropic 2025	→ Chinese notes · original
OpenAI Harness Engineering	Agent = Model + Harness / Humans steer, Agents execute	OpenAI 2026	→ Chinese notes · original

4. Glossary (83 entries × 10 topics)

Every HTML page embeds the full glossary in the lower-right corner. Click an orange marked term in the text and a popover opens without leaving the page. Grouped by topic below (click to jump):

Software engineering basics (10)

lint type check CI/CD sandbox canary A/B test HITL SLO/SLI/SLA TTL GMV / UV / CSAT / NPS / conversion rate

LLM fundamentals (9)

token embedding context window prefill/decode KV cache autoregressive logit temperature attention sink

Transformer architecture (5)

LayerNorm/RMSNorm RoPE YaRN/NTK GQA/MQA MoE multi-head attention

Inference acceleration (8)

FlashAttention PagedAttention vLLM continuous batching speculative decoding prefix caching TTFT TPOT

Quantization (4)

FP16/BF16/FP8 INT4/INT8/NF4 AWQ GPTQ

Fine-tuning and alignment (10)

SFT LoRA QLoRA RLHF RM DPO PPO GRPO RLVR catastrophic forgetting

Prompting and reasoning paradigms (7)

CoT ToT ReAct Reflexion Self-Refine Self-Consistency few-shot/zero-shot

RAG (9)

RAG chunking BM25 hybrid search reranker vector DB embed models RAG variants RAGAS

Agent engineering (11)

harness MCP tool use Computer Use sub-agent AGENTS.md deny-first progress file Loop of Death Memory Corruption LangChain

Evaluation and safety (10)

Eval LLM-as-judge eval benchmarks Trace + Eval tool stack Lost in the Middle prompt injection hallucination guardrails jailbreak

5. Suggested review paths

Path A · New graduate / beginner (2–4 weeks)

Week 1: Engineering foundation concepts — L1 §1.1–§1.3 + glossary (LLM basics + Transformer) + A1–A11 Week 2: Prompting + RAG introduction — L1 §1.4 + L2 §2.4 + C1–C9 + D1–D5 Week 3: Agent concepts — L3 §1–§3 + E1, E2, E3, E5, E8 + translation Building Effective Agents (Chinese notes) (original) Week 4: Safety + practice — L1 §1.6 + I1, I2 + J7 (0-to-1 rollout path)

Path B · Mid-level (3–6 years, 1–2 months)

Foundations (already in place) → deepen: • All of categories A–F • Read L1–L4 straight through • Parts of G/H (system design + evaluation) • Read all 4 translations • Prepare category J behavioral answers from your own projects

Path C · Senior / staff (depth + breadth)

• All 11 categories (A–K) • L4 §3.4 production-stack architecture + §3.4.6 maturity-model self-check • §3.5.6 eight cost-optimization levers • §3.6 scenario selection checklist • Memorize: Hashimoto's and Anthropic's own wording on harnesses • Have ready: 1–2 quantified project cases (GMV / conversion rate / accuracy, with numbers)

Path D · Scenario-driven (tailored to the target role)

Target role	Focus chapters
LLM application SaaS (customer support / document QA)	L2 RAG + L3 Agent + G1 (customer support) + I (safety) + K5 (business evaluation)
AI coding tools	All of L3 + L4 + G2 (Coding Agent) + E7 (harness) + translations of Effective Harnesses and OpenAI Harness Engineering
Multi-agent business systems	L3 §6 multi-agent + E4/E10 + K5 (business-domain evaluation) + G1/G3 + J1 STAR + numbers from your projects
Infrastructure / inference / training	A4–A9 inference acceleration + B1–B10 training stack + G4/G5 gateway and high concurrency
Security / red team	C7 + I1–I6 + L1 §1.6 + L4 three-layer defense

🧠 AI Agent / LLM application engineer: panoramic knowledge map