Beyond Sequential Agents: Orchestrating Hierarchical LLM-Based Agents for Complex Task Decomposition

Abstract

In the emerging field of LLM-based agent systems, coordinating multiple agents to solve complex tasks remains a significant challenge. This presentation demonstrates a practical implementation of hierarchical planning using Python and modern LLM APIs, drawing from real-world experience in building multi-agent task automation systems. By utilizing tools like LangChain, CrewAI, OpenAI’s API, and custom prompt engineering techniques, we showcase how to create a robust hierarchy of specialized agents that can decompose complex tasks, manage dependencies, and adapt to changing requirements. The talk delves into implementing manager-worker agent patterns, handling inter-agent communication through structured prompts, and maintaining coherent long-term planning across multiple agents.