Learn Agentic AI
2026An 8-day course for building AI agents in Python from first principles, graded deterministically, runnable offline, zero API costs.
The problem
Agent frameworks make everything feel like magic. You call a function, tokens stream out, tools somehow get invoked. I wanted to understand the machinery underneath: the tool-call loop, structured output parsing, retrieval pipelines, evals. And I wanted to practice it without burning API credits on every exercise run.
The key idea: a fake LLM you can grade against
The course's engine is llmlab, a deterministic fake LLM laboratory that mirrors the real SDK shapes (a FakeOpenAI, a FakeAnthropic, and MiniMCP for the Model Context Protocol). Exercises import it exactly like the real SDKs, but responses are deterministic, which makes them gradeable: the same input always produces the same output, so a test harness can assert on agent behavior precisely.
That one decision unlocks everything else: the course is free to run, works offline, and never flakes because a provider had a bad day. Swapping in a real SDK afterward is a one-line import change, because the shapes match.
What the 8 days cover
Provider adapters and streaming, structured output with validation and retries, the tool-calling loop written from scratch, MCP client and server, RAG with prompt-injection defense, a mini agent framework that ties it together, evals for measuring agent quality, and a FastAPI + SSE capstone that streams an agent's work to the browser.
The platform reuses the architecture I built for Express Academy and FastAPI Academy: a React/Vite client talks to a FastAPI grader that spawns each submission in an isolated child runner. Grader tests here are named Python snippets executed against your module, not HTTP specs, closer to how you'd actually unit-test agent code.
Outcome
Eight days, a 33-test quality gate passing end-to-end, and all three courses (Express, FastAPI, Agentic AI) run side-by-side on separate ports, a personal backend-to-AI curriculum that grades itself.