Course
AI Agents and MCP
Give a model tools and control what it can do with them: agent loop, MCP protocol, permissions, guardrails, and run observability.
Advanced 7 lessons
Lessons
- 1 A model that acts You'll be able to choose between a single call, a fixed chain, and an agent, and defend that choice by putting a number on the real cost of autonomy.
- 2 The agent loop You'll learn to write an agent loop with a reliable stop condition, an iteration budget, and history management that keeps your bill from exploding.
- 3 Describing tools You'll learn to write a tool schema that prevents bad calls and to return errors the model can correct instead of just suffering through them.
- 4 The MCP Protocol You'll be able to read the exchanges in an MCP session, tell tools, resources, and prompts apart, and choose a transport with confidence.
- 5 Writing an MCP server You'll implement a complete MCP server in Python and verify its behavior by hand, with no agent or model involved.
- 6 Permissions and guardrails You'll learn to separate read from write, require human confirmation, isolate files and network, and treat everything the agent reads as hostile.
- 7 Logging and Costs You'll learn how to record an agent trajectory in a usable way, reconstruct after the fact what was called and why, and set the limits that keep a loop from getting expensive.