Let's Data ScienceLEARN • BUILD • STAY AHEAD

LearnBuilding AI Agents

Building
AI Agents.

Build AI agents from first principles — the loop, tool use, planning, reflection, memory, multi-agent systems, evaluation, and guardrails. The model is a deterministic stand-in; every piece of agent machinery runs for real on the Helpwell support agent. Runnable Python, no API keys.

Course Overview

Modules

8

Duration

~9 hours

Helpwell Support AgentReal Loop, Faked LLMRunnable Python (Pyodide)ReAct → Memory → Multi-Agent

Module 1 is free. The full 8-module course is part of Pro.

Learning Modules

Each module combines animated explanations, hands-on Python practice, and a knowledge check. You build ONE thing — the Helpwell support agent — and watch it grow: from a bare loop to an agent that plans, reflects, remembers, collaborates, is measured, and is safe to ship.

Module 1: From Chatbot to Agent: The Loop

A chatbot talks; an agent acts. Build the agent loop from scratch — perceive, reason, act, observe, stop — plus the three parts (model, tools, instructions) and the stop rule that keeps it from running forever.

Sections

1Chatbot vs Agent — Who Owns the Plumbing

2The Agent Loop — Perceive, Reason, Act, Observe, Stop

3The Three Parts — Model, Tools, Instructions

4The Model Decides, Your Code Acts

5Why a Stop Rule Is Non-Negotiable

Module 2: Hands: Tool Use & Structured Outputs

Give the loop real hands. Function calling (the model requests, your code runs), schema-valid arguments so it never crashes on prose, ReAct (Thought → Action → Observation), and why the tool description is the biggest accuracy lever.

Sections

1The Tool-Use Contract — Name, Args, Id

2The Model Picks; Your Code Presses the Button

3Structured Outputs — Schema-Valid Arguments

4ReAct — Thought, Action, Observation

5The Tool Description Is the Biggest Lever

Module 3: Thinking Ahead: Planning & Decomposition

Intermediate Pro

Think before acting. Turn one fuzzy goal into an ordered to-do list (Plan-and-Solve), handle step dependencies, and re-plan when reality surprises you — "refund my last order AND cancel the renewal".

Sections

1What a Plan Is — One Goal, Ordered Steps

2Plan-and-Solve — Think First, Then Do

3Dependencies — Refund the Id You Just Found

4Plan, Observe, Re-Plan

5Decompose-First vs Decide-as-You-Go

Module 4: Second Drafts: Reflection & Self-Correction

Intermediate Pro

The first answer is a draft. Self-Refine, Reflexion, and CRITIC: a critic checks the draft against an objective rule, the agent revises — with the honest caveat that pure self-reflection often does nothing.

Sections

1The First Answer Is a Draft

2Self-Refine — Localized, Actionable Feedback

3Reflexion — Turn Failure Into a Written Lesson

4CRITIC — Check Against an External Rule

5The Evaluator–Optimizer Loop, and When to Stop

Module 5: Memory: Giving the Agent a Past

Intermediate Pro

A model is a consultant with amnesia. Working vs long-term memory, context rot, virtual memory (MemGPT), self-editing memory blocks, and scored recall (recency + importance + relevance).

Sections

1Statelessness — A Consultant With Amnesia

2The Finite Window and Context Rot

3Virtual Memory — RAM vs Disk

4Self-Editing Memory Blocks

5Scored Recall — Recency, Importance, Relevance

Module 6: A Team of Agents: Multi-Agent Systems

A team beats a soloist — sometimes. Handoffs, the supervisor/orchestrator pattern, the shared message bus, and the honest tradeoff: ~15x the cost and the single-writer rule.

Sections

1Handoffs — Transfer Control to a Specialist

2The Supervisor — Orchestrator and Workers

3Shared State — The Message Bus

4Roles, SOPs, and Debate

5The Honest Tradeoff — Cost and the Single-Writer Rule

Module 7: Does It Actually Work? Evaluation

Measure, do not eyeball. Outcome vs trajectory eval, task success rate, and pass^k reliability (does it succeed EVERY time, not just once?) on a gold set of support tasks.

Sections

1Measure, Don't Eyeball

2Outcome Eval — Grade the Final State

3Trajectory Eval — Grade the Path

4Task Success and pass^k Reliability

5LLM-as-Judge and Its Biases

Module 8: Trust & Ship: Guardrails, Safety & the 2026 Frontier

Trust and ship. Cost/loop caps, human-in-the-loop, prompt injection and the Lethal Trifecta, Meta's Rule of Two — and the 2026 frontier: MCP, computer use, and agentic RAG.

Sections

1Caps and Sandboxes — The Circuit Breaker

2Human-in-the-Loop for High-Impact Actions

3Prompt Injection and the Lethal Trifecta

4The Rule of Two

5The 2026 Frontier — MCP, Computer Use, Agentic RAG

Building AI Agents | Let's Data Science | Let's Data Science