AI agents, made clear.

The shortest honest path from “What is an agent?” to “I can ship one responsibly.”

An AI agent is not magic, a digital employee, or simply a chatbot with a dramatic name. It is software in which a model helps decide what to do next, within boundaries that people design. That modest definition unlocks the whole subject.

This course builds one idea at a time. You will first learn to recognize when an agent is useful. Then you will build a small, inspectable loop; add context, tools, memory, and approvals; evaluate it on realistic tasks; defend it against attacks; and operate it in production. The final capstones ask you to prove—not merely claim—that your system is useful and safe enough for its intended job.

What you will be able to do

By the end, you will be able to:

• explain agents without hype and distinguish them from models, chatbots, scripts, and workflows;
• choose the least complex architecture that can reliably solve a task;
• design tools, context, state, memory, permissions, and stopping conditions;
• build a bounded agent loop and a provider-backed implementation;
• use retrieval and Model Context Protocol connections without confusing access with trust;
• add approvals, budgets, retries, durable state, and human escalation;
• test behavior with reproducible evaluations instead of impressive demos;
• trace failures, contain incidents, and reason about latency and cost;
• build two portfolio capstones with evidence, citations, and production-readiness rubrics.

Choose your route

You do not need to read all 22 chapters before getting value.

If you are a product leader, take the 90-minute route before approving an agent project. If you write software, take the builder path. If you own reliability, security, governance, or a customer-facing deployment, take the production path.

The one mental model we will keep

Think of an agent system as Mission Control coordinating a flight:

The analogy has limits. A model is probabilistic, can be confidently wrong, and does not possess human intent. “Mission Control” is a thinking aid—not permission to anthropomorphize the software.

How the learning system works

Reading creates familiarity; it does not guarantee recall or transfer. This course uses techniques with stronger evidence:

• Retrieval practice: close the page and produce an answer from memory.
• Spacing: revisit ideas after delay rather than cramming once.
• Interleaving: distinguish neighboring concepts, such as state versus memory and workflow versus agent.
• Worked-example fading: watch one solution, complete a guided version, then solve independently.
• Prediction: commit to an answer before seeing the explanation.
• Concrete cases plus abstraction: understand one realistic system, then extract the reusable pattern.

Each chapter therefore follows the same rhythm: Predict → See → Explain → Practice → Retrieve → Revisit. A diagram is included only when a relationship is easier to understand visually. Decoration is not instruction.

The confidence rule

At every recall checkpoint, rate your answer from 0 to 4:

• 0: no answer;
• 1: guessed;
• 2: partly right;
• 3: correct but slow;
• 4: correct, quick, and explainable.

Low confidence and wrong answers are valuable signals. Review those first. High confidence plus a wrong answer is the most important signal of all: your mental model needs repair.

The course case: Northstar Travel

Examples build a fictional company called Northstar Travel. Its customers ask questions such as “Can I move my refundable booking to Friday?” A real answer may require policy retrieval, booking lookup, price calculation, a protected change operation, and sometimes a person.

The case grows with you:

1. a deterministic policy workflow;
2. a bounded tool-using assistant;
3. an evaluated research agent;
4. a customer-support agent with approvals and escalation;
5. a production service with traces, budgets, and incident controls.

No example uses real customer records, secrets, or irreversible actions.

Your safe local laboratory

The labs use a tiny standard-library Python system in courses/ai-agents/labs/. It is intentionally inspectable. A scripted model stands in for an external model so the core exercises run without an API key, network access, or charges. Later provider tracks show how to replace that test double with a real API.

From the Articulet project directory:

python3 -m unittest discover -s courses/ai-agents/labs/tests -v

Safety rules for every lab:

1. use synthetic data;
2. give tools the narrowest possible permissions;
3. separate read operations from write operations;
4. never place secrets in prompts, examples, traces, or source control;
5. simulate consequential actions until an approval boundary is tested;
6. cap steps, time, and spend;
7. treat retrieved or tool-returned text as untrusted data.

Provider-neutral core, dated implementation tracks

The enduring concepts in this course are provider-neutral. Chapters 11 and 12 add a dated OpenAI implementation track, checked against official documentation on July 17, 2026. APIs, model names, SDKs, framework features, regulations, and security guidance change. Dated material is labeled so a learner knows what to recheck.

Do not copy a 2026 model identifier into a future production system without reading the current provider documentation. Good agent engineering includes detecting which facts can expire.

What this course will not teach

It will not teach you to expose hidden chain-of-thought, give a model unrestricted system access, disguise automation as a human, bypass a site’s rules, or treat an agent’s confident output as proof. It will teach you to ask for concise explanations and evidence, log observable actions, test outcomes, and preserve accountable human control.

The graduation standard

Finishing pages is not graduation. You graduate when you can show:

• a task contract with explicit boundaries;
• an architecture justified against a simpler alternative;
• safe tools and approval points;
• a representative evaluation set and baseline;
• traces that make failures diagnosable;
• a release decision based on thresholds;
• a rollback and incident plan;
• honest documentation of residual risk.

Start with Chapter 1. Keep your answers. By Chapter 22, you will inspect your first definition of an agent and see exactly how your judgment changed.

Evidence behind the learning design

• Retrieval practice can outperform repeated studying and concept mapping (Karpicke & Blunt, 2011)
• Review of effective learning techniques (Dunlosky et al., 2013)
• Practice testing and distributed practice review paper