AI Agent

A regular chat AI is like a consultant: you ask, it answers, and you do the work. An agent is more like an assistant who does the work. You give it a goal, and it reads files, edits code, runs commands, and checks errors on its own, adjusting based on the results until it's done or needs you to decide something. Cursor's Agent mode and Claude Code both fall into this category.

It relies on Tool Calling. The model itself can only output text; tool calling lets it read and write files and run commands. Model Context Protocol (MCP) is a common standard for connecting external tools, so agents can use more of them.

Agents are powerful, but they may change a lot of files in one go. Give them a clear scope and a clear definition of done. Watch the commands they want to run, especially deleting files, installing dependencies, or changing the database. Save a snapshot with Git before changes.