RUNTIME EXECUTION ENVIRONMENTS

CHECK YOUR UNDERSTANDING • What is a sandbox and why do agents need one? ◦ Answer: an isolated place to run code that limits harm, needed because agents run code that may be buggy or malicious. • How do containers differ from virtual machines? ◦ Answer: containers share the host's operating system and are lighter and faster, while virtual machines are complete pretend computers with stronger but heavier isolation. • Why is an allowlist safer than a blocklist? ◦ Answer: attackers can use things a blocklist never listed, while an allowlist refuses everything not approved. • When is parallel execution appropriate? ◦ Answer: when steps are independent and rate limits, collisions, and cost allow it. • Name three parts of a graceful failure. ◦ Answer: any three of catching and classifying the error, a clear user message, logging, cleanup, and preserving partial results. PART FOUR: THE TOOLING LAYER (GIVING THE AGENT CAPABILITIES) The Execution layer taught you how the agent loop runs. The Tooling layer teaches you what the agent can actually do inside that loop. A model with no tools is a brilliant person locked in a room with no phone, no computer, and no door. It can think and talk, but it cannot look anything up, change anything, or help anyone beyond what it already knows. Tools are the windows and doors. In this Part you will learn what tools are (Chapter 13), how the harness catalogs them (Chapter 14), how to design them well (Chapter 15), how a shared standard called the Model Context Protocol connects them (Chapter 16), and the common families of tools you will meet in practice (Chapter 17).