Too Many AI Agent Skills Can Hurt Performance and Waste Tokens

A concept called 'skills sprawl' describes what happens when AI coding agents like Google Antigravity are overloaded with too many installed skills, leading to decision fatigue and reduced tool-selection accuracy. Agent skills work by loading descriptive metadata upfront and only pulling full skill content into context when needed, a mechanism known as progressive disclosure. While each skill's frontmatter uses fewer than 100 tokens, activating a skill can add thousands of tokens to a session, making excess skills costly. Having too many skills can confuse the model during task selection, increasing the chance of errors and inflating token usage per session. Pruning unused or redundant skills from an agent's configuration can improve reliability, reduce costs, and help the agent make faster, more accurate decisions.