Most "AI agents" are a single LLM wrapped in a system prompt. They look great in demos and collapse the moment real users show up. I know because I built one for my 5,000+ member developer community, and the first version broke in ways I didn't see coming.
In this talk, I'll walk through the multi-agent system I run in production: a tree of 12 specialist agents built on Google's ADK, coordinating through sequential pipelines, parallel fan-out, and LLM-driven dynamic routing, each armed with MCP tools to act on the real world. We'll start with the orchestration primitives, then climb into the real architecture: how the onboarding flow chains three agents in sequence, how external knowledge searches GitHub, Dev.to, and StackOverflow in parallel, and how the root agent decides which specialist to delegate to per message.
Then the fun part: the bugs. The drain-loop that cancelled a ParallelAgent mid-flight and surfaced the wrong reply. The recency drift that made the bot recommend 2020 articles in 2026. The callback layer I had to build for PII sanitization, response caching, and observability ; none of which made it into the tutorial.
You'll leave with a mental model for when to use each pattern, the production patterns that kept it running, and the war stories that'll save you a week.

Target audience: Intermediate developers familiar with LLMs and APIs. No prior agent-framework experience needed.
Preferred duration: 25 to 30 minutes. Can be tightened to 20 if needed.
Format: Conference talk with live walkthrough of the production architecture. Slides plus a short live demo of the system running locally (recorded backup available).
Key takeaways:

When to reach for each orchestration pattern: Sequential, Parallel, Dynamic routing
How to compose patterns into a real architecture (not one flat agent, but a tree)
Why MCP tools are the missing half of agent design: orchestration is the skeleton, MCP is the muscles
The callback layer nobody writes about: PII sanitization, response caching, lifecycle telemetry, date injection
Production war stories you can steal, with the fixes for each

Tech stack covered: Google Agent Development Kit (ADK), Model Context Protocol (MCP), Gemini, Python.
First public delivery. New talk based on a multi-agent system actively running in production.