Start here
What this course promises
Section titled “What this course promises”The goal is to understand how an LLM agent works as software: a model call inside a Python application that owns the tools, state, limits, and permission checks. We build that picture gradually, then ask when an agent or a second component is unnecessary.
This is an agent-focused curriculum. It does not teach model training or attempt a comprehensive survey of machine learning. No agent framework, workflow engine, tracing platform, or companion application is required.
What you should recognize
Section titled “What you should recognize”You should be comfortable reading basic Python functions, dictionaries, lists, loops, imports, exceptions, and simple type annotations. Familiarity with JSON helps when tool arguments move between JSON text and Python objects.
The first lesson introduces the model-call vocabulary. Lesson 2 introduces Pydantic models and enums. Later lessons explain the response items, JSONL logs, and handoff contracts they use. FastAPI appears as an occasional comparison, not a prerequisite. You do not need to have followed Max’s original chat or completed another course.
How the examples work
Section titled “How the examples work”The recurring situations are fictional customer support, invoices, document review, and company policies. IDs, addresses, amounts, policies, and sample outputs are synthetic teaching material.
Some code fences contain complete small examples. Others are focused excerpts that rely on definitions stated nearby, or conceptual pseudocode that isolates one decision. The lessons identify those boundaries. A repeated snippet lets you read its explanation without scrolling back; it is not necessarily another independent program.
Follow the transitions. Lesson 3 deliberately demonstrates one tool-request/result cycle. Lesson 4 introduces the general loop. Lesson 6 changes tool results to an ok/data/error envelope. Lesson 8 first returns an answer string, then changes the result to a dictionary so an evaluator can inspect the answer and requested tools. Lesson 9 adds an explicit question argument so the writer receives the original request.
Reading requires no installation, account, or API key. If you elect to run API examples, you will need your own configured credentials and model access, and inference has a cost. The examples are not a production service to clone and deploy.
Read in order, then return by topic
Section titled “Read in order, then return by topic”For a first reading, follow lessons 1 through 10. Each group builds on the preceding one:
- Lessons 1–4: the model-call boundary, structured output, tools, and the application-owned loop.
- Lessons 5–7: state and context, failure handling, and permission and approval boundaries.
- Lessons 8–10: evidence about behavior, cooperation through contracts, and architectural judgment.
For a return visit, use the sidebar, the page’s contents list, or search. Search includes code as well as prose, so a name such as approve_deletion or ResearchBrief can bring you back to an explanation.
Technical assumptions and verification
Section titled “Technical assumptions and verification”The reviewed Python examples use Python 3.13.18, OpenAI Python SDK 3.55.1, and Pydantic 2.14.4 where structured parsing appears. API examples consistently use the documented gpt-5.5-2026-04-23 snapshot. The technical review date is November 19, 2026. Model availability and APIs can change; consult the linked primary documentation when adapting an example.
No reasoning-effort or output-token override is set in these examples. Timeout, retry, and step limits have different meanings, which the reliability lessons explain. None is presented as a universal guarantee of cost or elapsed time.
The checks use extracted Python, fixed fixtures, and synthetic responses through the installed SDK. They verify selected application behavior and interfaces. They do not establish live model quality, account access, production authentication, durable execution, or safe concurrency. Sample outputs, timings, scores, and token counts are illustrative unless explicitly identified as measured.
Throughout the course, keep three distinctions clear: a valid schema can contain a wrong decision; a model request does not grant permission; a polished answer is not proof that an action happened or a source supports the prose.