Agentic Engineering & ToolingDeep Dive180min
Agent Loops, Decoded: Fundamentals to Agentic Patterns in Spring AI
This talk introduces Spring AI agents as context orchestrators, starting with ChatClient, memory, Advisors, RAG, Tool Calling, and MCP. It then builds advanced agent patterns: tool discovery, evaluation loops, explainable agents, planning, user questioning, subagents, and agent-to-agent integration, previewing Spring AI 2.1’s typed multi-step agent abstraction.
talk.summaryAiDisclaimer
Christian TzolovSpring Team @ Broadcom
talkDetail.whenAndWhere
Monday, October 5, 09:30-12:30
Room 5
At their core, AI agents are context orchestrators — gathering information, querying models, evaluating outputs, and adapting their approach. This live-coded session takes you from Spring AI fundamentals to agentic patterns.
We start with the fundamentals: ChatClient essentials, context engineering, and session memory. We cover how Advisors extend the LLM execution path, including Recursive Advisors implementing the Agent Loop — the loop behind every autonomous agent — then enrich context with RAG and connect to external services via Tool Calling and MCP.
From there, we build up the agentic patterns Tool Calling and Advisors make possible:
You'll leave understanding how a few building blocks compose into any agentic pattern, with an early preview of what's next: durable, crash-recoverable agentic workflows in Spring AI 2.1.
We start with the fundamentals: ChatClient essentials, context engineering, and session memory. We cover how Advisors extend the LLM execution path, including Recursive Advisors implementing the Agent Loop — the loop behind every autonomous agent — then enrich context with RAG and connect to external services via Tool Calling and MCP.
From there, we build up the agentic patterns Tool Calling and Advisors make possible:
- Tool Search Tool — dynamic tool discovery for agents with hundreds of tools
- LLM-as-a-Judge — automated response evaluation in an evaluator-optimizer loop
- Explainable Agents — augmenting tool calls with the model's reasoning for debuggability
- Agent Skills — LLM-agnostic capabilities, loaded on demand
- Todo / structured planning — preventing "lost in the middle" failures
- AskUserQuestion — gathering requirements before acting
- Subagent Orchestration — delegating to specialized agents with isolated context
- MCP, A2A, and ACP — securing tool access, building rich MCP Apps, and connecting agents to other agents and clients
You'll leave understanding how a few building blocks compose into any agentic pattern, with an early preview of what's next: durable, crash-recoverable agentic workflows in Spring AI 2.1.