Summary
In this chapter, we explored how to enhance LLMs by integrating tools and design patterns for tool invocation, including the ReACT pattern. We started by building a ReACT agent from scratch and then demonstrated how to create a customized one with just one line of code using LangGraph.
Next, we delved into advanced techniques for controlled generation—showing how to force an LLM to call any tool or a specific one, and instructing it to return responses in structured formats (such as JSON, enums, or Pydantic models). In that context, we covered LangChain’s with_structured_output method, which transforms your data structure into a tool schema, prompts the model to call the tool, parses the output, and compiles it into a corresponding Pydantic instance.
Finally, we built our first plan-and-solve agent with LangGraph, applying all the concepts we’ve learned so far: tool calling, ReACT, structured outputs, and more. In the next chapter, we’ll continue...
