The Spring AI Ecosystem in 2026: From Foundations to Agents
Mark Pollack - Broadcom / Christian Tzolov - Broadcom
In 2026, the question for Java teams is no longer “how do I call an LLM?” The industry has moved up the stack—toward agentic workflows where AI reasons about goals, plans multi-step tasks, and takes action. The new question is: how do I harness this effectively?
This session maps the entire Spring AI ecosystem, from foundation to frontier.
We begin with Spring AI 2.0—ChatClient, advisors, tool calling, structured outputs, memory, and vector stores. These building blocks let you create sophisticated AI-powered applications with fine-grained control. But in 2026, this is the foundation, not the destination.
Next, Model Context Protocol (MCP). We’ll cover the latest spec updates, MCP Annotations for declarative @McpTool development, and MCP Security for enterprise authentication and compliance. MCP defines how models access tools, resources, and context—and it’s matured significantly.
Before climbing the agentic stack, we need to understand what makes agents tick. At their core, AI agents are context orchestrators—continuously gathering information, querying models, evaluating outputs, and adapting their approach. Spring AI’s Advisor architecture enables these behaviors by acting as middleware in the AI interaction pipeline, where capabilities like Memory and RAG become composable building blocks. But real-world agent workflows demand more than linear processing: they need iterative tool invocation, validation with automatic retry logic, and continuous quality assessment. We’ll introduce Recursive Advisors as a mechanism for controlled iteration, with hands-on demonstrations showing what it takes to build your own MemGPT or Claude Code-like agentic systems.
With that foundation, we climb the agentic stack.. Over twenty agentic CLI and IDE tools now exist—Claude Code, Gemini CLI, Amazon Q, Amp, Codex, and more. The Spring AI Agent Clients project provides a unified interface to invoke any of them from Java. The Claude Code SDK for Java offers formal, fully-typed integration. Need more control? Spring AI Agent Harness project helps you build custom agentic loops, paired with a curated set of tools and skills designed specifically for agents. The Judge framework ties it together—verifying that agents achieved their goals, whether you’re evaluating off-the-shelf agents or your own agentic loop implementations.
Finally, emerging protocols: ACP (Agent Client Protocol) standardizes client-to-agent communication—Gemini CLI supports it, and you can build ACP-compliant agents that plug into IDEs. A2A (Agent-to-Agent) enables direct agent collaboration.
Whether you’re building chatbots or deploying autonomous coding agents, you’ll leave with a map of the entire Spring AI ecosystem and practical guidance on where to start in this constantly evolving landscape.