Building AI Agents in 21 Days: A Hands-On Course in Agentic Systems with Python and LLMs (Learn in 21 Days)

Building AI Agents in 21 Days
A Hands-On Course in Agentic Systems with Python and LLMs
by M. Saqib

================================================================
Go from a single line that calls a language model to a deployed,
autonomous agent that plans, uses tools, checks its own work, and
recovers from failure — in three focused weeks.
================================================================

Most "AI agent" books hand you a framework and a magic import. You
end up with a demo you can't debug and only a hazy idea of what the
agent actually does. This book is the opposite. You build the agent
yourself, from the loop up, in plain Python — so when something
breaks, you know exactly why.

You call Claude through the official Anthropic Python SDK and write
every part by hand: the control loop, the tools, memory, retrieval,
planning, guardrails, and the deployment around it. Frameworks like
LangGraph and CrewAI are explained so you can read them — but you
never hide behind one. By Day 21 an agent is no longer magic. It is
something you can build, measure, and ship.

WHAT YOU GET
- 21 day-chapters (3 weeks x 7 days) of careful, worked teaching —
no filler, no hand-waving.
- 150+ runnable Python listings with real output.
- 246 figures and diagrams — loop diagrams, sequence diagrams, and
measured plots.
- A clean, print-friendly layout, and a per-day workshop (quiz and
exercises) so each idea lands in your hands.

THE 21 DAYS
Week 1 — Foundations
Day 1. What Is an Agent? The Loop That Thinks
Day 2. Tool Calling — Giving the Model Hands
Day 3. The Agent Loop
Day 4. ReAct — Reasoning Before Acting
Day 5. Memory & the Context Window
Day 6. Structured Output You Can Trust
Day 7. Errors, Timeouts, and a Resilient Loop
Week 2 — Capability
Day 8. Designing Tools the Model Can Use
Day 9. Retrieval as a Tool (RAG)
Day 10. Planning vs. Reacting
Day 11. Multi-Step Tasks & Working State
Day 12. Self-Correction & Reflection
Day 13. Cost and Latency Budgeting
Day 14. Tracing and Observability
Week 3 — Robustness and Scale
Day 15. Evaluating Agents
Day 16. Guardrails and Safety
Day 17. Human in the Loop
Day 18. Multi-Agent Orchestration
Day 19. Long-Running & Background Agents
Day 20. Deploying an Agent
Day 21. Capstone — A Complete Agent, End to End

WHO IT'S FOR
You know ordinary Python — functions, dictionaries, exceptions. You
do NOT need any machine-learning background, and you will not train a
model; you call one and build the system around it. You need an
Anthropic API key; each call costs a fraction of a cent, and the
whole book runs for about the price of a coffee.

BY DAY 21 YOU CAN
- build the agent loop that calls a model, runs its tools, and feeds
the results back until the task is done;
- design tools, schemas, and structured output the model uses well;
- ground an agent's answers in your own documents with retrieval;
- make an agent plan, reflect, and correct itself;
- measure and cut what every run costs in tokens and time;
- evaluate an agent like code, guard it against bad input and prompt
injection, and deploy it as a real service.

The agent stops being a black box. Go build one that works.