A loop with a model in it — and every guard that keeps it from running away.
7 lessons · about 7 hours · free
An agent calls tools in a loop, deciding after each result whether it is finished. That is the whole idea, and this course builds it: the loop itself, state carried between steps, and breaking a task into steps that can be checked. Then the half that matters in production — step limits so it terminates, budgets so it cannot spend without bound, permissioning so a model cannot delete what it was only meant to read, sandboxing so a mistake is contained, and a human-in-the-loop gate in front of anything irreversible. It also covers MCP, the protocol that lets an agent discover tools it was not hard-coded to know about.
What makes something an agent, and how state survives between steps.
Limits, budgets, permissioning, sandboxing, and the gate in front of anything irreversible.