Topic 2
Three kinds of tool
AI Engineering (Ch. 6, "Tools") says a system doesn't need external tools to be an agent, but without them its capabilities are limited: a language model by itself can only generate text. The set of tools an agent can use is its tool inventory, and the book groups tools into three kinds.
Predict first
Scroll sideways to see the whole diagram
Knowledge augmentation: tools that build context. Retrievers and SQL executors from Module 05, plus web search so the model doesn't go stale.
Text version
- Knowledge augmentation: tools that build context. Retrievers and SQL executors from Module 05, plus web search so the model doesn't go stale.
- Capability extension: tools that cover a model's weak spots. AI Engineering's example: a model will likely fail 199,999 ÷ 292, a calculator won't.
- Write actions: tools that change something. They let a system do far more, from replying to email to moving money.
- The book's comparison: you wouldn't let an intern delete your production database, and you shouldn't let an unreliable AI initiate bank transfers.
- Knowledge augmentation: tools that build context. Text and image retrievers, SQL executors, internal people search, an inventory API, an email reader, and web browsing, which keeps a model from going stale.
- Capability extension: tools that cover what models are bad at. The book's example: a model will likely fail at 199,999 divided by 292, but the calculation is trivial with a calculator. Also calendars, unit converters, translators, and code interpreters, which the book notes come with the risk of code injection attacks.
- Write actions: tools that change the environment. A SQL executor that can delete a table, an email API that can reply, a banking API that can transfer money.
Read-only tools let an agent perceive; write tools let it act. The book's caution on writes: just as you shouldn't let an intern delete your production database, you shouldn't let an unreliable AI initiate bank transfers. Lesson 6 is about how to grant write access safely.
More tools give an agent more capabilities, but the book also notes the more tools there are, the harder it is to understand and use them well. Choosing them is an experiment, like everything else.
Quick check