A conference talk by IBM engineers Mario Fusco and Kevin Dubois covering agentic AI patterns implemented in LangChain4J. They walk through building multi-agent systems using four core workflow patterns: sequence (chaining agents one after another), loop (iterating until quality threshold is met), parallel (concurrent agent execution), and routing (conditional agent dispatch). The talk introduces a pluggable Planner abstraction that lets developers implement custom orchestration patterns and compose them with existing ones. Key concepts include shared agentic scope for cross-agent context, agent monitoring with HTML report generation, a supervisor pattern where an LLM dynamically decides which agents to call, and a goal-oriented pattern using graph-based shortest-path planning. Live demos show a story generation pipeline, a medical/legal/technical question router, and a currency transfer supervisor. The framework is part of LangChain4J's agentic module, actively releasing monthly updates.
Sort: