Build Your Own AI Agent — the Loop, the Tools, and MCP (Season 1)

Build Your Own AI Agent — the Loop, the Tools, and MCP (Season 1)

Uncategorized
Wishlist Share

About Course

You have used an AI agent. Now build one.

In this course we write the agent loop by hand, in Python, with no framework. We give it
tools, let it touch a real machine behind a permission layer, teach it to recover when things
go wrong, put it on MCP so anything can talk to it, manage its context and cost, and finally
write evals that tell you whether a change made it better or worse.

By the end you will have built Rover: a working terminal coding agent, and the judgement to
know what it should and should not be trusted with.

What you will learn

  • Write an agent loop from scratch, with no framework
  • Design tools the model chooses correctly, and debug it when it does not
  • Give an agent safe access to files and a shell behind a permission layer
  • Recover from runaway loops, false finishes, and every stop reason that matters
  • Build an MCP server and an MCP client, and explain the protocol underneath
  • Manage context, caching, and memory so a long session stays sharp
  • Write evals that tell you whether a change helped

Requirements

  • Comfortable in a terminal, and able to read and write code in some language
  • Python 3.10 or newer, and an editor
  • An Anthropic API key (a small amount of credit covers the whole course)
  • No prior experience with agents or MCP

Who this is for

  • Developers who have used coding agents and want to know how they work
  • Anyone building an internal tool on top of an LLM and hitting the limits
  • Engineers who need to review or maintain agent code somebody else wrote
Show More

Course Content

Module 1 — An agent is a loop
The whole mechanism, in one module. Everything after this is detail.

Module 2 — Tools: teaching the model what it can do
The part every tutorial skips. Tool choice is a writing problem before it is a code problem.

Module 3 — Letting it touch your machine
Where a toy becomes something you have to be careful with.

Module 4 — When the loop goes wrong
Recovery cannot be taught in prose. This module is video-first.

Module 5 — MCP: the protocol under the tools
The reason this course is worth taking now. Almost nobody teaches the mechanism.

Module 6 — Context: the resource you are actually managing
Why the agent felt sharp for twenty minutes and useless after an hour.

Module 7 — Proving it works, and what it costs
Ends with the learner's own agent, not ours.

Student Ratings & Reviews

No Review Yet
No Review Yet