MEPX
Chapter 7 of 10All chapters

Chapter 7 of 10

Tools and agents

Letting a model act.

Adding capability

Giving a model tools, such as search, code execution or an interface to your systems, extends what it can do and multiplies what a mistake can cost.

  • An agent loop can repeat an error many times before anyone notices.
  • Permissions, review steps and limits matter more as autonomy grows.

Prompt injection

A model reading a web page or document will treat instructions inside it as text worth following. Anything acting on untrusted content needs a boundary between data and commands.