Building AI Agents from First Principles

You’ve probably seen AI agents in action—reasoning through complex problems, orchestrating API calls, recovering from mistakes, and maintaining intent across dozens of interactions. But how do they actually work? This workshop answers that question from first principles. Through a series of progressively more capable examples, we’ll evolve a simple request/response interaction into a complete agentic system with an explicit control loop: plan → act → observe → decide. Rather than treating agents as black boxes, we’ll expose the architectural patterns that make them reliable, understandable, and testable.

The examples are implemented with Spring Boot and Spring AI, but this is not a framework tutorial. The focus is on the fundamental patterns that underpin agentic systems, regardless of language, framework, or AI model. Using a dinner planning assistant as our running example, we’ll progressively introduce tool invocation, planning, memory, multi-step execution, evaluation, error handling, and termination conditions, showing how each capability emerges naturally as the agent evolves.

A central theme of the workshop is that agents are systems embedded in systems. Building robust agents is less about prompt engineering and more about software architecture. By the end of the session, you’ll have a clear mental model of the agentic loop and a practical set of architectural patterns you can apply to build production-ready agentic systems, regardless of the technologies you use.