Chapter 9: LLM Agents & Tools

Language models on their own can only generate text. But when equipped with tools -- APIs, databases, code interpreters, web browsers -- they become agents capable of taking actions in the world. An LLM agent observes its environment, reasons about what to do, selects and executes tools, and iterates until the task is complete.

The key insight behind LLM agents is that language models are remarkably good at reasoning about when and how to use tools, even though they were trained only on text. Given a description of available tools and a task, an LLM can decompose the problem, select appropriate tools, format correct API calls, interpret results, and decide what to do next.

Tool use transforms LLMs from knowledge retrieval systems into action-taking systems. Instead of just answering "the weather in Tokyo is typically warm in July," an agent can call a weather API and report "it's currently 28C and sunny in Tokyo." Instead of explaining how to query a database, it can write and execute the SQL.

The ReAct (Reasoning + Acting) framework formalizes this into a loop: Thought (reason about the situation), Action (call a tool), Observation (interpret the result), repeat. This interleaving of reasoning and action enables complex multi-step problem solving.

Multi-agent systems take this further by having specialized agents collaborate. One agent might handle research, another coding, another review -- mimicking how human teams work.

This chapter covers:

• Tool Use & Function Calling: How LLMs invoke external APIs and functions with structured arguments
• ReAct Pattern: The Thought-Action-Observation loop for systematic problem solving
• Multi-Agent Systems: Orchestrating multiple specialized agents for complex tasks
• Memory & Context: Short-term and long-term memory architectures for persistent agents
• Agent Frameworks: LangChain, LlamaIndex, CrewAI, and other tools for building agents