An agent is a model that can do more than answer: it calls tools such as search, code execution, file access or an API, reads the results, and decides what to do next. Every tool description, every tool result and every instruction becomes part of the agent's context, so the way tools are designed directly affects how well the agent performs.
The lessons cover the basic agent loop of environment, tools and prompt, then the Model Context Protocol (MCP), an open standard for connecting models to tools and data sources. You will also see agent skills, which package expertise in files that an agent loads only when needed, so the context window stays small. The earlier modules on memory and retrieval show where this information can come from, and this one shows how an agent reaches for it.
Watch the lessons in order, then continue on to Testing, Evals & Prompt Safety.
Barry Zhang of Anthropic on when an agent is worth building, how to keep it to an environment, a set of tools and a system prompt, and why you should think from the agent's limited point of view.
Video 15minAnthropic's Mahesh Murag on what MCP is, how to build with it, and how agents use MCP servers to reach tools, data and prompts, ending with where the protocol is heading.
Video 1h 44minBarry Zhang and Mahesh Murag on agent skills: folders of instructions and scripts that an agent loads on demand, giving it domain expertise without filling the context window.
Video 16min