Agent

Agent is a broad word, which is part of the problem. In AI tooling, it usually means a model wrapped with tools and instructions so it can do more than answer a prompt.

An agent might read files, call APIs, search docs, edit code, run tests, or open a pull request. The useful bit is the loop: observe, decide, act, inspect the result, then continue.

The risk is also in that loop. An agent with tools can make real changes. Good guard rails matter: clear instructions, scoped permissions, readable logs, and a human review before anything important ships.