Agentic architectures
As we learned in Chapter 5, agents help humans solve tasks. Building an agent involves balancing two elements. On one side, it’s very similar to application development in the sense that you’re combining APIs (including calling foundational models) with production-ready quality. On the other side, you’re helping LLMs think and solve a task.
As we discussed in Chapter 5, agents don’t have a specific algorithm to follow. We give an LLM partial control over the execution flow, but to guide it, we use various tricks that help us as humans to reason, solve tasks, and think clearly. We should not assume that an LLM can magically figure everything out itself; at the current stage, we should guide it by creating reasoning workflows. Let’s recall the ReACT agent we learned about in Chapter 5, an example of a tool-calling pattern:
Figure 6.1: A prebuilt REACT workflow on LangGraph
Let’s look at a few relatively...
