In depth
An AI agent wraps an LLM in a loop: perceive → plan → act → observe → repeat. The LLM is the reasoning engine; tools are its limbs; memory (short-term context + long-term stores) is its continuity. The agent's autonomy comes from its ability to choose which tool to call next based on the goal and what it has learned so far.
Agents range from simple (single-step tool use) to complex (multi-agent orchestration with sub-agents, playbooks, and self-reflection). Claude Code, Cursor, and Windsurf are agentic coding tools — they plan a change, edit files, run tests, and iterate on failures autonomously.
Modern agents use MCP as their tool protocol. The LLM sees a list of MCP tools (from connected servers), picks one, calls it, reads the result, and decides the next step. This decouples the reasoning (LLM) from the capabilities (MCP servers).
Agents are characterized by three properties: **autonomy** (decide next action), **tool use** (call external capabilities), and **memory** (remember context across turns). Add multi-agent orchestration and you have the current frontier.