Building Intelligent Agents
As generative AI adoption grows, we start using LLMs for more open and complex tasks that require knowledge about fresh events or interaction with the world. This is what is generally called agentic applications. We’ll define what an agent is later in this chapter, but you’ve likely seen the phrase circulating in the media: 2025 is the year of agentic AI. For example, in a recently introduced RE-Bench benchmark that consists of complex open-ended tasks, AI agents outperform humans in some settings (for example, with a thinking budget of 30 minutes) or on some specific class of tasks (like writing Triton kernels).
To understand how these agentic capabilities are built in practice, we’ll start by discussing tool calling with LLMs and how it is implemented on LangChain. We’ll look in detail at the ReACT pattern, and how LLMs can use tools to interact with the external environment and improve their performance on specific tasks...
