How does an Agent decide which Tool to use next?

• A. By rolling a die.
• B. By consulting the current plan and tool metadata, aided by the LLM's reasoning to select the most appropriate tool.
• C. By always using the first tool in the registry.
• D. By asking the user for permission before each step.

Answer: (B)

The decision hinges on a plan-driven approach that uses tool metadata to pick the best fit. The agent has a plan that outlines the next objective and the information needed to reach it. It also maintains a registry of tools, each with metadata describing what the tool can do, what inputs it expects, what outputs it returns, and any constraints or limitations. When it’s time to act, the LLM reasons about which tool best matches the required action given the current inputs and state, considering compatibility, data types, and any constraints (like latency or permissions). The tool that most closely fits the next step is chosen and invoked, and the result updates the plan and state for the following actions. This makes tool use purposeful and adaptable, rather than random, and avoids fixed or user-permission-only behavior.