There is no universally agreed definition of what an AI agent is. In practice though, several patterns are emerging. These patterns demonstrate the coordination and integration of multiple AI services to build sophisticated Agentic AI systems capable of handling intricate tasks.

These Agentic Systems architectures can be grouped in 2 main categories: workflows, where LLMs and tools are orchestrated through predefined code paths, and autonomous agents, where LLMs dynamically direct their own processes and tool usage, maintaining control over how they execute tasks.

The goal of this talk is to give a theoretical overview of Agentic AI in general and these patterns in particular. We will discuss their differences and range of applicability and show with practical examples how they can be easily implemented and tested. We’ll use Quarkus and its LangChain4j extension, but the concepts are universal.

Devoxx

A conference talk by Kevin Dubois and Mario (IBM/Red Hat, Quarkus team) covering agentic AI patterns implemented in the LangChain4j agentic module. The talk walks through four core orchestration patterns: sequence (chaining agents), loop (iterative refinement with exit conditions), parallel (concurrent agent invocation), and conditional routing (directing queries to specialized agents). It introduces the concept of 'agentic scope' as a shared state across agents. Beyond rigid workflows, a supervisor pattern is shown where an LLM autonomously decides which agents to invoke and in what order. Finally, a pluggable planner SPI is introduced, enabling custom patterns like goal-oriented (A* search) agent graphs. Recent additions include streaming chat model support and multimodal agents. Live demos illustrate routing, shared context enrichment, and a money transfer scenario using the supervisor pattern.

Agentic AI Patterns by Kevin Dubois