Watching lessons only gets you so far. These five builds take the ideas from the modules and make you apply them: you'll wire up an agent loop, expose your own tools over MCP, give an agent memory, assemble a research agent with a framework, and then attack an agent to see where it breaks.
Each project follows along with a video walkthrough. Work through them in order, or jump to the one that matches the module you just finished.
Wire up a plain LLM API call into a working agent loop with a couple of custom tools — no framework, just to see what's actually happening under the hood.
Create a custom Model Context Protocol server exposing tools and resources, then connect it to a real agent host and test it end to end.
Use a framework to assemble a multi-tool agent that searches the web, reads pages, and summarizes findings into a report.
Attempt prompt injection and jailbreak attacks against a working agent using free CTF-style labs, then apply what you learn to sandboxing, input validation, and tighter tool permissions.