How LLM agents act in the world: tool use, the ReAct-style agent loop, planning and memory, MCP, multi-agent systems, and why agents are hard to get right.