Building Multi-Agent AI Systems on Databricks

Welcome to July!

We are excited to invite you to our July meetup, which will be taking place online.

This month, we will be joined by Paul Karikari for a practical workshop on “Building Multi-Agent AI Systems on Databricks: Patterns & Pitfalls”.

Multi-agent AI systems are becoming increasingly popular, but many examples still stop at impressive demos rather than showing how these systems can be built, traced, evaluated, and governed in real environments. For teams working with data, AI, and enterprise systems, the bigger question is not just whether multiple agents can work together, but when they are genuinely useful, how to design them sensibly, and what can go wrong when they move beyond the prototype stage.

This workshop will give attendees hands-on experience with building multi-agent AI systems on Databricks. The session will introduce four core design patterns that appear in many production systems; it will also show how these patterns map onto Databricks-native tools such as Mosaic AI Agent Framework, MLflow, Vector Search, and Unity Catalog. You’ll be building a multi-agent prototype from the ground up in a provided Databricks notebook.

We will begin by exploring when a single-agent system is enough and when a multi-agent approach is worth considering. We will then work through the main design patterns before moving into a hands-on lab where attendees build a supervisor-style research agent, connect specialist components, trace the system with MLflow, and run an offline evaluation.

The session will also cover common pitfalls that real teams encounter when building multi-agent systems, including unnecessary complexity, unclear agent responsibilities, weak evaluation, and governance gaps. Attendees will leave with a practical mental model, a completed notebook, and a checklist they can use when designing or reviewing multi-agent systems in future.

No prior experience with Databricks is required, although you should be comfortable with Python and have a basic understanding of LLMs or prompt-based applications.