# Agents Mental Model A simplified mental model for understanding how [[AI Agents]] work, broken down into five layers. ## Agent loop → Reasoning The core loop (observe → think → act → evaluate) is the agent's reasoning engine. It decides what to do next, breaks down tasks, recovers from errors, and iterates until the goal is met. ## Runtime → Execution The runtime provides the agent's hands. Bash, filesystem access, code execution, and other tools let the agent act on the world. Without a runtime, the agent can only talk. ## MCP servers → Connections [[Model Context Protocol (MCP)]] servers extend the agent's reach beyond its local environment. Each MCP server is a connection to an external system (databases, APIs, SaaS tools, browsers), giving the agent access to capabilities it doesn't have natively. ## Skills → Institutional memory Skills are reusable instructions that persist across sessions. They encode how to do things (workflows, conventions, best practices) so the agent doesn't have to rediscover them every time. Skills are institutional memory that actually persists. ## Skills library → Domain expertise A collection of skills forms a skills library, which gives the agent domain expertise. A well-curated skills library turns a general-purpose agent into a specialist (e.g., a frontend developer, a DevOps engineer, a technical writer). ## Related - [[AI Agents]] - [[Model Context Protocol (MCP)]] - [[Claude Code]]