Lecture 5 · COMS W4995 Agentic Engineering

Agent tools in pictures

Lecture 5 · October 9, 2026

Agent tools in pictures

What a tool is, how a call happens, five ways to hand one to an agent, and the judgment calls that come with it.

Nick GuCOMS W4995 Agentic Engineering · Columbia University · Fall 2026

This book goes with Lecture 5, “What tools does an agent actually need?” It starts from the smallest thing that is true about a tool call and builds up from there, one idea at a time.

Each section ends with a question to answer before you look. Each chapter ends with a quiz where most questions have more than one right answer. Every fact links to its source at the end of its chapter.

Contents

Read the opening and chapter 1 before class
  1. OpeningA model writes the order, a program cooks it7 min · 4 questionsBy the end of these pages you can say why a model never runs a tool itself, name the four questions every tool raises, and tell what is forty years old about tool calling and what is new.
  2. Chapter oneForty years of calling someone else's code22 min · 8 questionsFollow remote calls from 1984 to 2026 and you can tell which parts of today's tool protocols are lasting problems and which are packaging that will change again.
  3. Chapter twoWhat a tool is22 min · 10 questionsTake any tool apart into the parts the model sees, the parts that run, and the parts nobody wrote down, then ask it four questions before you trust it.
  4. Chapter threeHow a call happens22 min · 10 questionsFollow one get_weather call from the definition you write, through the tokens the model sees, to the harness that runs it, and learn what each step does and does not guarantee.
  5. Chapter fourFive ways to hand a tool to an agent28 min · 10 questionsThe same capability can reach an agent as a function, a shell command, a skill, an MCP server or code; by the end you can say where each one keeps its parts, what it buys you, and when it buys you nothing.
  6. Chapter fiveMany tools22 min · 10 questionsWhy more tools can hurt, and what to try, in order, before you hand the model another one.
  7. Chapter sixThe trust boundary32 min · 10 questionsDraw one map of everyone between a request and its effect, then use it to decide who must say yes to each call, whose words may steer it, and whose key it carries.
  8. Chapter sevenTools in an organization12 min · 6 questionsWhen a whole company connects agents to its systems, someone has to decide who may use which tool, as whom, and through which door; by the end you can say where that control lives and what it does not cover.
  9. Chapter eightMeasuring tool use20 min · 8 questionsLearn to say where a run failed, to read a benchmark number together with the setup that produced it, and to test your own tool server with a small experiment instead of a single score.
  10. Chapter nineThe judgment calls20 min · 8 questionsFive decisions shape every tool server, and you can test each one by asking whether it would still matter with a model ten times better.
  11. Chapter tenWhat is still open14 min · 5 questionsFive questions the field has not settled; for each, what the sources establish, what nobody has published, why it matters for a tool you build now, and what evidence would settle it.