Integrating Tools and Selection Strategy

Every LLM, regardless of capability, has a fundamental limitation: it is a closed system. If we ask an AI chatbot, “What’s the current weather in San Francisco?” the model must respond with something like, “I cannot access real-time weather information.” This demonstrates that an LLM, on its own, operates as a closed system. It has no access to live internet data and no ability to perform real-world actions. In essence, it functions entirely within its internal model state, without external awareness or interaction.

However, modern LLMs with agentic capabilities are explicitly designed to overcome this limitation. They are built with the latent ability to use external tools. This could mean calling a weather API, searching a private database, or interacting with other software. This ability to take action is what transforms a simple chatbot into a capable AI agent.

This new capability raises the critical question: How does the model know what tools are available to it? How does it understand the purpose of each tool? And how does it decide which one to use for a specific user request? This is where prompt engineering is essential. Our role as prompt engineers is not to build these agentic capabilities, but to direct them. Through the prompt, we provide the playbook or the user manual that the AI follows to take action in a controlled and predictable way.

Let’s explore the prompt engineering techniques for defining, selecting, and planning with external tools. We will learn how to write effective tool definitions that serve as a guide for the model and how to utilize advanced prompting frameworks, such as ReAct, to orchestrate complex, multi-step tasks.

The fundamentals of tool use

To instruct a model to use tools, we must first understand the structure of the information that needs to be provided in the prompt. This process, often referred to as function calling or tool use ...