Modelli di progettazione dell'IA generativa

Chapter 7. Enabling Agents to Take Action

The patterns we’ve covered in this book so far have involved creating content in some form. In this chapter, we’ll discuss three patterns that enable applications built on foundational models to interact with the world—often, this is the line beyond which we consider the application to be agentic. Tool Calling (Pattern 21) enables models to invoke external functions through structured interfaces. Code Execution (Pattern 22) allows models to write and run code to solve complex problems. Multiagent Collaboration (Pattern 23) orchestrates specialized AI agents in hierarchical, peer-to-peer, or market-based architectures to tackle complex tasks through division of labor.

Pattern 21: Tool Calling

The Tool Calling pattern enables an LLM to act in the world, either to obtain information (for example, through web search) or to make a change to the environment (for example, by writing out a file). It does so by having the LLM emit special tokens when it determines that a function needs to be called, along with the arguments to pass to that function. A client-side postprocessor invokes the function with those parameters and sends the results back to the LLM, and the LLM incorporates the function results in its response.

Problem

At their core, multimodal LLMs generate content in different forms (text, images, audio, and video). This is enough to provide capabilities such as creating research reports, translating between languages, and generating ...