Multi-agent architectures
In Chapter 5, we learned that decomposing a complex task into simpler subtasks typically increases LLM performance. We built a plan-and-solve agent that goes a step further than CoT and encourages the LLM to generate a plan and follow it. To a certain extent, this architecture was a multi-agent one since the research agent (which was responsible for generating and following the plan) invoked another agent that focused on a different type of task – solving very specific tasks with provided tools. Multi-agentic workflows orchestrate multiple agents, allowing them to enhance each other and at the same time keep agents modular (which makes it easier to test and reuse them).
We will look into a few core agentic architectures in the remainder of this chapter, and introduce some important LangGraph interfaces (such as streaming details and handoffs) that are useful to develop agents. If you’re interested, you can find more examples and tutorials...
