From Assistants to Agents: Self-Improving Agentic Systems with Spring AI
Christian Tzolov - Broadcom
AI agents are systems dedicated to the art of context window curation — relentless loops of assembling context, prompting the model, observing results, and re-assembling for the next step. Building them is more than stitching together LLM calls; it demands systems that can iteratively refine their context, self-correct, and act autonomously toward better outcomes.
This talk explores how Spring AI provides an elegant framework for exactly that. The ChatClient, Advisors, and Tool Calling APIs serve as the core building blocks: after briefly covering the essentials — prompt engineering, conversation memory, RAG for context enrichment, Tools for external services, and MCP for standardized integration — we’ll focus on the agentic patterns that turn a reactive AI assistant into something that can reason, plan, and act.
We’ll explore patterns such as Recursive Advisors for controlled iteration through the advisor chain (enabling output validation with retry, self-organizing memory, and custom guardrails); Agent Skills for modular, LLM-agnostic capabilities loaded on demand; Todo to prevent “lost in the middle” failures with structured planning; Subagent to delegate to specialized agents with isolated context windows; A2A and ACP Protocols for interoperable agents crossing system boundaries; LLM-as-a-Judge for automated response evaluation; and Tool Search for dynamic tool discovery.
Through live demonstrations using Spring AI’s fluent API, you’ll see how complex agentic behaviors reduce to just a few lines of clean, maintainable code. You’ll leave with a clear understanding of how to integrate generative AI into Spring applications — from your first prompt to a complete agentic architecture.