Level 3: Agents & MCP — Briefing
Agents use tools to solve tasks autonomously. MCP standardizes the connection to tool servers. In this level you’ll learn both — from the first tool call through multi-step agents to the standardized protocol for external services.
Skill Tree
Section titled “Skill Tree”
What You’ll Learn
Section titled “What You’ll Learn”- Tool Calling — How an LLM calls functions to perform actions instead of just generating text
- Tools in the Frontend — How to make tool calls visible in the UI with message parts and loading states
- Tool Loop Agent — Multi-step agents that autonomously use multiple tools in sequence
- MCP (Model Context Protocol) — A standardized protocol for connecting to external tool servers
- Tool Approval — Human-in-the-loop for critical operations, so agents don’t act without oversight
Why This Matters
Section titled “Why This Matters”Without tools, an LLM can only talk. It has no knowledge of current data, can’t read files, can’t call APIs, can’t compute anything. Tools turn a chatbot into an agent that gets things done.
The concrete problem: You want your LLM to perform real tasks — check the weather, fetch data from a database, run calculations. For that it needs tools. And if you want to connect many different external services, you need a protocol for that — MCP.
Prerequisites
Section titled “Prerequisites”- Level 1 completed —
generateText,streamText, system prompts, Zod schemas - Zod fundamentals — Schema definition,
.describe()for parameters - async/await — Tools are asynchronous functions
Skip hint: You already know
tool(),stopWhen: stepCountIs()andcreateMCPClient? Jump straight to the Boss Fight and build a research agent with MCP and approval flow.
Challenges
Section titled “Challenges”Boss Fight
Section titled “Boss Fight”Build a Research Agent — an agent that uses MCP tools, researches across multiple steps, and asks the user for approval before saving results. You’ll combine all five building blocks of this level in one project.