What does the term Excessive Agency signify in LLM applications?

• A. LLMs lacking any autonomy
• B. LLMs gaining the ability to make decisions or perform actions via extensions based on prompts
• C. LLMs relying solely on user input without automation
• D. LLMs disabled from making any decisions

Answer: (B)

Excessive Agency describes a situation where a language model isn’t limited to producing text but can act autonomously by using extensions to make decisions or perform tasks based on prompts. This means the model can interact with external tools, APIs, or systems, carry out actions, and influence real-world outcomes without always needing step-by-step human intervention. The idea is that the model’s autonomy goes beyond generating responses to actually executing tasks, which raises safety and governance concerns because actions can have unintended or harmful consequences if not properly controlled. For example, enabling plugins or tool use that let the model fetch data, run code, or control devices demonstrates this level of autonomy. The other descriptions describe levels of autonomy that are more constrained: no autonomy, strictly user-driven automation, or being unable to make decisions, none of which capture the model acting through extensions and making its own choices.