$ learn --agentic-ai --from-zero

A single, ordered path from "what is a token" to a deployed agent with memory, tools and evals — nothing skipped, nothing you have to guess the order of.

8 phases · 16 weeks · ~10 hrs/week

start phase 00

before you start

[ok] comfortable writing and debugging Python
[ok] has called a REST API before (curl, Postman, requests — any of them)
[ok] knows what a token and a context window are, roughly

the roadmap

Eight phases, in order. Each one builds on what came before it — do not skip ahead.

Foundations

// 2 wk · ~10 hrs/wk

tokens & context windows → embeddings, in plain terms → prompt design patterns → a working Python + SDK setup

Prompting & Structured Output

// 2 wk · ~10 hrs/wk

function calling → JSON-mode / structured responses → few-shot & chain-of-thought → scoring a prompt's output

Retrieval (RAG)

// 2 wk · ~12 hrs/wk

vector databases → chunking strategies → embedding + retrieval pipelines → a working document Q&A

Tool Use

// 2 wk · ~10 hrs/wk

designing a tool schema → calling real APIs from a model → retries & error handling → chaining multiple tools

Agent Frameworks

// 2 wk · ~12 hrs/wk

LangGraph state machines → LlamaIndex agents → CrewAI multi-agent crews → picking the right one for the job

Memory & State

// 2 wk · ~8 hrs/wk

short vs. long-term memory → summarizing long conversations → persistent stores (SQLite / Redis) → session & user state

Evals & Observability

// 2 wk · ~10 hrs/wk

tracing with LangSmith → building an eval set → catching regressions → watching cost & latency

Capstone

// 2 wk · ~8 hrs/wk

scoping a real agent → build, memory, tools together → evals running in CI → ship a public demo

time budget

Same numbers as the roadmap above, laid out so you can put it on a calendar.

phase	weeks	hrs/wk	total hrs
00 Foundations	2	10	20
01 Prompting & Structured Output	2	10	40
02 Retrieval (RAG)	2	12	64
03 Tool Use	2	10	84
04 Agent Frameworks	2	12	108
05 Memory & State	2	8	124
06 Evals & Observability	2	10	144
07 Capstone	2	8	160
total	16	~10	160

the stack

What you will actually have installed and calling by the end of phase 04.

[provider] OpenAI / Anthropic — the model your agent actually calls

[orchestrate] LangGraph — stateful, multi-step agent graphs

[retrieve] LlamaIndex — document loading, indexing, retrieval

[coordinate] CrewAI — multi-agent crews with defined roles

[store] a vector DB — Chroma, pgvector, or Pinecone — pick one

[observe] LangSmith — tracing, evals, regression checks

definition of done

By the end of phase 07 you have a deployed agent that holds a real conversation across sessions, calls at least two tools of its own choosing, and has an eval suite that runs on every change — not a notebook, a running thing with a URL someone else can open.

Phase 00 starts today.

begin phase 00