The Art of Building Agents
Adib Saikali - Broadcom
You’ve probably seen AI agents in action—reasoning through messy problems, orchestrating API calls, recovering from mistakes, and somehow maintaining intent across dozens of steps. At some point you’ve wondered: how does this actually work—and could I build one?
This full-day workshop answers those questions from first principles. Through a series of progressively complex examples, you’ll see how a simple request/response interaction evolves into a full agentic system with an explicit loop: plan → act → observe → decide. Rather than treating agents as black boxes, we’ll make the loop visible and connect it directly to the architectural decisions that determine reliability, safety, and testability.
A central theme of this workshop is that agents are systems embedded in systems. We move beyond prompt design to explore how API design, tool boundaries, and control flow shape what agents are able to do—and how safely they can do it.
This workshop is hands-on and exploratory. Bring your laptop and a willingness to learn. We’ll provide a GitHub repository with prepared examples and API keys for AI services used in the exercises. By the end of the day, you’ll have a clear mental model of agentic systems—and practical architectural patterns you can apply to start building your own.
Key Concepts Covered
- How agents plan by reasoning about goals, creating action plans, and deciding what to do next
- How agents act in the world by interacting with APIs and external tools
- How agents evaluate progress—and why determining “am I getting closer to the goal?” is a hard problem
- How agents manage the context window across multiple steps, including short- and long-term memory
- How to design agentic systems where architecture, API design, and tool boundaries shape safe and reliable behavior
- How agents act securely on behalf of users—and where the risks lie
Hands-on Exercises
- Building intuition by experimenting with existing agents before writing code
- Reviewing key Spring AI concepts such as tool calling and advisors
- Building MCP (Model Context Protocol) servers and clients using Spring AI
- Constructing an explicit agentic loop from first principles
- Evolving a simple chat interaction into a reasoning agent
- Integrating tools and external capabilities using MCP
- Designing AI-friendly APIs using hypermedia and server-driven constraints (Spring HATEOAS)
- Managing context and state across multi-step interactions
- Adding termination conditions and error handling for robust behavior
- Applying agentic security mechanisms, including requesting and enforcing permissions using existing security standards
Prerequisites
- Comfortable with Java and Spring Boot
- Familiar with foundational AI concepts such as models, prompts, embeddings, and tool calling
- Basic experience with Spring AI (key concepts will be reviewed as needed)
- Interest in learning how to architect and build AI agents