A 12-week, project-driven curriculum for cloud / infrastructure engineers transitioning into AI Agent / LLM Engineer roles. Authored as an Obsidian vault. Companion to the lab repo shaneliuyx/agent-prep.
Tone: each chapter is a long-form pedagogical document — theory primer, hands-on lab, bad-case journal, interview soundbites. No tutorial graveyard. Every week ends with a runnable artifact + measured number.
Open Agent Development 3-Month Curriculum.md first — top-level overview + discipline rules. Then walk the weeks in order. Decimal-numbered weeks (Week 2.5, Week 2.7, etc.) are electives that cut between sequentially-prerequisite weeks; they're optional but strongly recommended for senior interview prep.
Week 0 — Environment Setup
Week 0.3 — Agent History + Foundational Narrative ← 5-era lineage (symbolic → reactive → BDI → RL → LLM) + 3-property agent def (hello-agents Ch1+2, SPEC, v0)
Week 0.5 — LLM Internals Speedrun ← tokenize → embed → QKV → sample (interview Q13/Q14 cover, SPEC, v0)
Week 1 — Vector Retrieval Baseline
Week 2 — Rerank and Context Compression
Week 2.5 — GraphRAG on a Wikipedia Subset ← v12.4m: 0.96 judge, 32/0/0 vs vector
Week 2.7 — Structure-Aware RAG (PageIndex) ← 16/16 GT-judge vs Vector 0.500 / Graph 0.375
Week 3 — RAG Evaluation ← RAGAS, HyDE, Phoenix
Week 3.5 — Cross-Session Memory ← 15/15 + Phase 5 mem0 cross-check (10/14)
Week 3.5.5 — Multi-Agent Shared Memory ← guild MCP integration + atomic-claim race
Week 3.5.5.5 — Multi-Agent Topology Patterns ← supervisor / hierarchical / group-chat / handoffs / voting (Anthropic 90.2% measurement)
Week 3.5.8 — Two-Tier Memory Architecture ← guild operational + EverCore semantic + Phase 8.6 bitemporal supersede/coexist (5/5 tests)
Week 3.5.9 — Requirement-Driven Memory Arch. ← LongMemEval decomposition + 6-backend matrix (1-tier / 2-tier / hybrid / three-tier HyperMem L3)
Week 3.5.95 — Self-Observability Memory ← PAI v7.6 OBSERVABILITY + LEARNING self-facing axes + Phase 7 heat-scored eviction (BAI-LAB/MemoryOS leverage; importance-exempt) — lab executed, tests 6/6
Week 3.5.96 — Self-Wiring Memory (GBrain) ← markdown-first deterministic graph + hybrid retrieval (measured: pure-vector > RRF on 19-page corpus — 83→95 is corpus-dependent, not universal) + Ground-Truth Hierarchy A/B (ClaudioDrews/memory-os) (Garry Tan / YC, MIT, lab executed)
Week 3.7 — Agentic RAG ← measured: canonical skip-allowed graph mis-built (faithfulness 0.876 vs single-pass 0.980, 15/50 skips) → structural always-retrieve = 1.000 at parity; CRAG web fallback decomposes comparisons (per-sub rerank) over a SearXNG backend, 10/10 out-of-corpus; web_search promoted to shared (lab executed)
Week 4 — ReAct From Scratch ← + §1.5 MLX Studio gateway role-map (2026-05-15 sync)
Week 4.5 — Model Routing and Effort Tiering ← local Qwen-1.5B classifier + cost-latency Pareto front (SPEC, v0)
Week 4.6 — Durable Agent Runtime + Topologies ← AutoGPT executor/scheduler/lock kernel + PraisonAI 4-mode process (SPEC, v0)
Week 5 — Pattern Zoo ← ReAct vs PaS vs Reflexion
Week 5.5 — Metacognition
Week 5.6 — ISA-Driven Metacognition ← PAI v5.0 Ideal-State Artifact as falsifiable termination contract (SPEC, v0)
Week 6 — Claude Code Source Dive
Week 6.4 — Low-Code Agent Platforms (Dify / Coze / n8n / LangFlow) ← 5-axis low-code-vs-custom matrix + Dify/Coze/n8n side-by-side (hello-agents Ch5, SPEC, v0)
Week 6.5 — Hermes Agent Hands-On
Week 6.6 — MCP Schema Bridge ← type-hint → JSON Schema producer + AsyncGenerator streaming (SPEC, v0)
Week 6.65 — MCP Production Transports ← Streamable HTTP + DNS-rebinding via Origin + HTTP+SSE 2026-migration + gateway composition + tool-poisoning catalog (SPEC, v0)
Week 6.7 — Authoring Agent Skills ← extended with microsoft/apm (Agent Package Manager, May 2026)
Week 6.75 — Skill Optimization (SkillOpt) ← microsoft/SkillOpt arXiv:2605.23904; text-space training of skills (epochs/batchsize/lr) without weight updates (SPEC+lab v0)
Week 6.8 — Agent Communication Protocol Survey ← MCP / A2A / ANP (SPEC, planned)
Week 6.85 — Prompt Template Engineering Patterns ← 5-axis design space + schema-enforce + 4 anti-patterns (interview Q1, SPEC, v0)
Week 6.9 — Context Engineering + Todo Mechanisms ← 4 context shapes + cognitive-narrowing argument + 80-LOC TodoList primitive (interview Q11, SPEC, v0)
Week 6.95 — A2A Protocol ← Agent Card + Task lifecycle + AP2 signed cards + MCP-vs-A2A decision matrix + Linux Foundation v1.0 April 2026 (SPEC, v0)
Week 7 — Tool Harness
Week 7.3 — Production LLM Infrastructure ← gateway + caching + cost attribution + fallback (Akshay 6-area #2+#5)
Week 7.5 — Computer Use and Browser Agents
Week 7.6 — Small-Model Agent Stacks ← MagenticLite + MagenticBrain (Qwen3-8B) + Fara1.5-9B local-first agent stack on M5 Pro (SPEC+lab v0)
Week 7.7 — Quantization + Inference Optimization ← FP16 → MXFP4 hierarchy + KV-cache math + memory-bound regime (JD#1, SPEC, v0)
Week 7.8 — Code-Agent Patterns (AST + Coverage + Mocks) ← tree-sitter + LSP + branch coverage + 4-class testability filter (interview Q3-Q7, SPEC, v0)
Week 8 — Schema Reliability Bench
Week 8.5 — Voice AI Agents
Week 8.7 — Generative Media + Fine-tuning ← diffusion + LoRA + ControlNet + IP-Adapter for brand-consistent gen (JD#4, SPEC, v0)
Week 9 — Faithfulness Checker
Week 9.3 — Agent Performance Evaluation ← AgentBench / GAIA / SWE-bench + 5-dim per-trajectory rubric + LLM-as-judge with Cohen's κ (hello-agents Ch12, SPEC, v0)
Week 9.5 — Agentic RL Fine-Tuning (SFT + GRPO) ← LoRA r=16 SFT + GRPO group-relative + 4-row ablation table on GSM8K-mini (hello-agents Ch11)
Week 10 — Framework Shootout ← LangGraph vs LlamaIndex vs OAI Agents SDK
Week 11 — System Design
Week 11.5 — Agent Security
Week 11.55 — Content Provenance + AI Regulatory ← SynthID + Stable Signature + C2PA layered defense + EU AI Act / US EO / UK / Korea + Model Cards + EchoLeak CVE (SPEC, v0)
Week 11.6 — Production Tracing + Cost Telemetry ← OpenTelemetry + Langfuse + DuckDB rollups + p99 vs mean argument + cost-governors + A/B canary (JD#2, SPEC, v0)
Week 11.7 — Take-Home Dress Rehearsal ← 4-hour timed take-home + evals-first git discipline + 5-min Loom defense + 30/30/25/15 rubric scoring
Week 11.8 — Continuous Training + MLOps ← PSI drift detector + eval-gated CI + shadow ramp + MLOps Level 3 (JD#3, SPEC, v0)
Week 12 — Capstone and Mocks
Week 12.5 — Multimodal Agents ← CLIP/BLIP-2/LLaVA substrates + ColPali vision-native RAG + cross-modal fusion + omni-models thinker-talker (SPEC, v0)
Interview-prep spine: see [[Interview Question Index]] for question → chapter → measured-anchor lookups. Use it as the pre-interview review entry point.
FDE career track: see [[FDE Track - Forward Deployed Engineer Path]] — an overlay note that maps every week above onto the Forward Deployed Engineer skill stack (agentic orchestration, evals and guardrails, production deployment, MCP integration, customer context) and gives a compressed FDE reading path. Read it if you are targeting Forward Deployed Engineer roles specifically. The customer-context gap it identifies is closed by Week 12's FDE Delivery Mode.
| Area | Where covered |
|---|---|
| 1. Harness engineering | W4, W5, W7, W7.8 (code-agent variant) |
| 2. Inference serving (KV cache, paged attention, spec decoding, quantization) | W0, W0.5 (LLM internals primer), W2.7 BCJ #23, W7.7 (quantization + KV-math + memory-bound regime), W9.5 |
| 3. Structured output reliability | W8, W6.85 (prompt-template schema-enforce layer) |
| 4. Evals + observability | W3, W2.7, W3.5, W11.6 (production tracing + cost telemetry), W11.8 (CT eval-gated CI) |
| 5. Production LLM infrastructure (gateway, caching, cost attribution, fallback) | W7.3, W11.6 (OTel + Langfuse + DuckDB cost rollups), W4 §1.5 (MLX Studio gateway) |
| 6. Fine-tune vs in-context | W9, W9.5, W8.7 (diffusion LoRA + DreamBooth for brand-consistent generation) |
| 7. Production ML / CI-CD-CT (additional JD axis) | W11.8 (PSI drift detection + eval gates + shadow ramp + MLOps Level 3) |
| 8. Code-agent skills (additional JD axis) | W7.8 (AST + LSP + branch coverage + mock-injection + 4-class testability filter) |
The following common interview-Q clusters map to dedicated chapters:
| Question pattern | Chapter |
|---|---|
| Q1: "How are prompt templates constructed?" | W6.85 |
| Q3-Q7: AST / coverage / mocks / code parsing | W7.8 |
| Q9: "Parallelized intent recognition" | W4.5 §Phase 4 vote layer |
| Q11: "Context engineering / todo lists" | W6.9 |
| Q12: "Skills implementation" | W6.7 |
| Q13: "What does the model see when you type X?" | W0.5 |
| Q14: "Self-attention / QKV mechanics" | W0.5 |
See Agent Development 3-Month Curriculum.md → "The Six Areas an AI Engineer Must Master" for the mapping table. Trend-monitoring cadence (~30 min / week) keeping this current is documented in Trend-Monitoring Discipline.md.
- Not a tutorial collection. Tutorials teach syntax. This teaches measured engineering — every claim is anchored in numbers from a lab
RESULTS.mdin theagent-prepsibling repo. - Not framework-evangelism. Each framework week (LangGraph, LlamaIndex, OpenAI Agents SDK) ends with a head-to-head comparison on the same task, so readers can pick on data, not vibes.
- Not opinion-free. Each chapter takes a position (when GraphRAG wins, when HyDE loses, when reasoning models are an anti-pattern), defends it with measurements, and shows the alternative cases.
- Hardware: MacBook Pro M5 Pro, 48 GB unified memory (M-series Apple Silicon).
- Local-first inference: oMLX serving Gemma-4-26B-heretic / Qwen3.6-35B-A3B / gpt-oss-20b on
:8000, plusbge-m3-mlx-fp16for embeddings. - Vector DB: Qdrant via OrbStack (Docker) on
:6333. - Memory infra (Weeks 3.5.5 / 3.5.8 / 3.5.9):
mathomhaus/guild(Go MCP, single binary, embedded SQLite) for operational tier; EverMind-AI's EverCore (Python + Postgres via Docker compose, port 1995) for semantic tier; HyperMem (Docker compose, port 1996) for relational L3 tier. Benchmarked via LongMemEvaloraclesubset anchored to EverCore's published 83%. - Observability: Phoenix on
:6006. - Cloud spend cap: ~$13 across the program (W7–8 frontier-model comparisons ~$8 + W7.3 cross-provider gateway routing ~$3 + W9.5 optional cloud GPU $0–30). Diagnostic threshold $20 — exceed it, audit which lab is leaking. Used only when local can't substitute.
Lab code: shaneliuyx/agent-prep — every chapter's lab phases are runnable artifacts there.
Every chapter follows the same 10-section template, defined in CLAUDE.md:
- Why This Week Matters (~150 words — production motivation + interview signal)
- Theory Primer (~1000 words — papers cited, real numbers)
- Mechanism / Architecture Diagram (mermaid, every node labeled)
- Lab Phases (numbered, time-budgeted, with per-Python-block bundle: Architecture mermaid → Code → Walkthrough → Result → ★ Insight)
- (deprecated — walkthroughs live inline next to their code per §4 per-block-bundle rule as of 2026-05-07)
- Bad-Case Journal (3-5 entries: symptom / root cause / fix; mark
(observed YYYY-MM-DD)for entries from real lab runs,(pre-scoped)for theory-derived predictions pending validation) - Interview Soundbites (2-3 entries, ~70 words each, anchored to measurements)
- References (peer-reviewed papers + production blogs)
- Cross-References (Builds on / Distinguish from / Connects to / Foreshadows — bidirectional invariant)
- Frontmatter (YAML — title, created, updated, tags, audience, stack)
Week 2 - Rerank and Context Compression.md is the canonical reference; new chapters should match it structurally.
MIT. Curriculum prose is original; references in each chapter cite their original authors (Anthropic, Microsoft GraphRAG, RAGAS, RAPTOR, PageIndex, etc.).