Advanced Overview
What Advanced covers, what you'll build by the end, and how it builds on the multi-tool agent you already shipped in Intermediate.
Chapter 1: Multi-Agent Patterns
Supervisor, hierarchical, and swarm patterns for splitting work across more than one agent -- and a lab that builds a supervisor by wrapping two specialist agents as tools.
Chapter 2: Advanced RAG
Query rewriting, HyDE, multi-hop retrieval, and self-correcting RAG -- four techniques for when a single vector search over the raw question isn't enough, tried against the same corpus Chapter 3 used to break plain vector search.
Chapter 3: Fine-tuning vs. RAG vs. Prompting
When fine-tuning actually earns its cost over prompting and RAG, plus a real (tiny) LoRA fine-tune you train and compare yourself, no GPU, no API key, a few seconds on a laptop CPU.
Chapter 4: Guardrails and Safety
A hand-rolled guardrail wrapper: pattern-based input checks against prompt injection, Pydantic schema validation on the output, retry-then-fail-closed -- and an honest look at where that minimum version actually breaks.
Chapter 5: Observability
Instrumenting an agent's model calls with OpenLLMetry (OpenTelemetry) so every call becomes a span with the prompt, the completion, the token counts, and the timing -- exported to your own console, no account required.
Chapter 6: Production Concerns
Three things a demo never needs but real traffic always does: caching repeated questions, rate limiting bursts, and streaming output -- implemented from scratch, no library, so you see exactly what each one is actually doing.
Chapter 7: Shipping It
Wrapping the support bot in a small FastAPI app, then writing a Dockerfile to package it so it can run anywhere, not just on your machine -- including the one networking gotcha every Docker beginner hits.
Chapter 8: Capstone — A Guarded, Traced, Evaluated Agent
The Advanced capstone — the Intermediate capstone's three-tool agent wrapped in an input guardrail and full OpenLLMetry tracing, then measured with precision@k/recall@k and LLM-as-judge against the whole agent, not just retrieval.