Most "AI agents" are a single LLM wrapped in a system prompt. They work in demos, then collapse under real-world complexity. The fix isn't a bigger model; it's specialized agents that coordinate, each armed with MCP tools to act on the real world.
In this talk, I'll live-code a multi-agent system using Google's Agent Development Kit (ADK) and MCP servers in Python. You'll see three orchestration primitives: SequentialAgent (pipeline), ParallelAgent (concurrent fan-out), and LoopAgent (iterative refinement). But orchestration alone isn't enough; agents that can only reason aren't useful. So each specialist gets MCP tools: a researcher agent searches the web, an analyst reads files through a filesystem MCP server, and a reviewer uses MCP-connected services to validate output.
The key insight: ADK gives you deterministic control flow between agents, MCP gives each agent its capabilities. Orchestration is the skeleton; MCP tools are the muscles.
You'll leave with working code, a mental model for when to use each orchestration pattern, and practical patterns for plugging MCP servers into multi-agent workflows that are composable, testable, and production-ready.