How AI agents work: an AI that can only write, and still gets things done

The model inside an AI agent can only write text. It has no hands and no internet. Still, AI agents check the weather, move meetings and send messages for you every day. A plain program sits next to the model and does what its text asks. Meet that program and the parts around it: tool calls, tool descriptions, the agent loop, permissions, MCP, the context window and subagents. Then see two ways it goes wrong: an agent that repeats the same step, and a web page with hidden orders in it.