MCP & Multi-Agent Systems
By Wanderson Leandro de Oliveira
This course is the third stop in the AI Engineering track, and it moves from a single agent to the world it actually operates in: standardized tool access via the Model Context Protocol (MCP), the frameworks that orchestrate multi-step workflows, the humans who need a say before risky actions happen, and the other agents an agent increasingly has to talk to. You will learn MCP's architecture and lifecycle, then build a real server and client with the official SDK. A full module is dedicated to MCP security from the builder's side — not how to exploit a server, but how to spot tool poisoning and path traversal in your own code and fix them, including a hands-on lesson where you patch a real vulnerable MCP server from the DARE Vulnerable AI Suite and verify the exploit stops working. From there the course covers choosing an agent framework without vendor lock-in regret, workflow-first design that mixes deterministic steps with agentic ones, durable execution that survives crashes, human-in-the-loop approval gates and escalation, patterns for coordinating multiple specialized agents, and the message contracts and trust model that let agents talk to each other safely. It closes with a capstone: designing a supervisor that delegates to two specialist agents with a defined message schema and a clear rule for when to call a human.
Course content
MCP Fundamentals
- 🔒 MCP architecture: hosts, clients and servers text
- 🔒 Connection lifecycle, transports and capability discovery text
- 🔒 Hands-on: building an MCP server and client text
MCP Security
- 🔒 Tool poisoning and supply chain risk text
- 🔒 Defending MCP servers text
- 🔒 Hands-on: fixing a vulnerable MCP server text
Frameworks and Workflows
- 🔒 Choosing an agent framework text
- 🔒 Workflow-first design text
- 🔒 Durable execution and resumable workflows text
Human-in-the-Loop
- 🔒 Approval gates and risk classification text
- 🔒 Edit-before-execute and confirmation UX text
- 🔒 Escalation and fallback text
Multi-Agent Systems
- 🔒 Specialization and delegation text
- 🔒 Coordination patterns text
- 🔒 Failure modes in multi-agent systems text
Agent-to-Agent Communication
- 🔒 Message schemas and contracts text
- 🔒 Trust, identity and authorization between agents text
- 🔒 Capstone: a supervisor with two specialist agents text