Lecture 17: LLM Agents & Tool Use

From prompting to acting: reasoning, tools, and multi-agent systems

Overview

Large Language Models are more than chatbots—they can plan, call tools, and take actions. This lecture shows how to turn an LLM into an agent: structuring outputs, invoking external functions/APIs, and iterating with reasoning-and-action loops (ReAct). We contrast naive prompting with constrained and schema-validated approaches, calibrate model decisions via log-prob scoring, and discuss evaluation, safety, and failure modes when models act in the real world.

Learning Objectives

By the end of this lecture, you will be able to:

• Understand LLM versatility - Recognize that LLMs are general-purpose text processors capable of classification, extraction, reasoning, and generation tasks
• Implement tool use - Build systems that enable LLMs to interact with external functions and APIs through structured tool calling
• Create ReAct agents - Develop agents that combine reasoning and acting in an iterative loop to solve multi-step problems
• Fine-tune models - Customize language models for specific domains through efficient fine-tuning techniques

Materials

TipQuick Access

LLM Agents & Tool Use Notebook

Resources

• Classic Blog Posts:
  • A Visual Guide to LLM Agents — Architectures for reasoning, planning, and tool use

  • LLM Powered Autonomous Agents

Previous: ← Lecture 16: Recurrent Neural Networks | Next: Lecture 18: Coming Soon →