Lesson 25 — Capstone: a study-buddy agent

Everything together

Your study buddy uses every idea from this course:

• a guardrail that refuses unsafe requests before the model sees them;
• long-term memory (memory) that is sent to the model inside the system message;
• a model that replies in JSON: remember, a final answer, or a plan of tool calls;
• three tools — calculator, clock and glossary — run from a registry;
• five eval cases that must all pass.

Two pieces are missing. When the model says {"remember": {"name": "Dana"}}, save it into memory (memory.update(...) adds a dict's keys to another dict). When it sends a plan, run every step and collect the results as text.

Why it matters: this is the real shape of an agent. Swap fake_llm for a model API call and it is a working app — the evals tell you whether the swap kept it working.