COMMON TOOL CATEGORIES

CHECK YOUR UNDERSTANDING • Which tool category is generally the most dangerous and why? ◦ Answer: code execution and shell, because code can delete files, leak secrets, or consume resources. • Why separate draft_email from send_email? ◦ Answer: so drafting can never accidentally send, and sending can have its own approval and limits. • Why is web content treated as untrusted? ◦ Answer: it can be wrong or contain hidden instructions that attempt prompt injection. • How can a database tool be made safer? ◦ Answer: read-only accounts, table limits, row limits, parameterized queries, time limits, and logging. • What is the minimum viable tool set principle? ◦ Answer: give the agent only the tools its purpose requires, to reduce cost, confusion, and attack surface. PART FIVE: THE CONTEXT LAYER (MANAGING THE AGENT'S WORKING MEMORY) Welcome to Part Five. In the earlier parts of this book, you met the model, the harness, and the loop that lets an agent think and act step by step. Now we turn to one of the most important jobs a harness performs, which is deciding what the model gets to see at each moment. Everything the model can use to think about your task has to fit inside one limited space, called the context window. Learning to manage that space well is what separates an agent that works for five minutes from an agent that works reliably for five hours. In this Part, you will learn why context is scarce, how it goes bad, how to build it carefully, how to shrink it safely, and how to store information outside of it.