Skills Guide, Plugins Enable: Why AI Agent Tools Need Real Executable Motors

A developer and researcher argues that AI agent ecosystems are misusing terms like skill, plugin, and tool as if they were interchangeable, when they serve distinct roles. A skill should teach an agent how to use a capability, while a plugin must actually make that capability available through executable components. The author identifies a problem called "context debt," where agents accumulate excessive Markdown instructions that consume model attention before any real work begins. Using a memory plugin as a case study, the piece shows how even well-architected plugins can fail when activation instructions double as dispatch orders rather than simply enabling capabilities. The proposed solution distinguishes clearly between skills as focused guidance, plugins as distribution units, and tools as bounded operations with inputs, outputs, and verifiable evidence.