This path is for you if a chat that only answers is not the tool in front of you. Parts separate chat, an agent, work mode, and a coding agent, then how much you let it do, how loops run away, and how to check the work without being an engineer. Start with the first part on this page.
- 1 What is an AI agent? A plain-English guide An agent plans, acts, observes, and acts again with tools you grant. Chat answers once. Write a stop, keep delete off, and watch the first run so a tidy folder does not erase the archive.
- 2 Chat vs agent vs work mode vs coding agent Chat, agent, work mode, and coding agent are four doors. Pick by the artifact. A short rewrite does not need a folder-wide agent run.
- 3 Autonomy levels: watch, approve, walk away Pick watch, approve, or walk away on purpose. If you typed save and the agent sent mail, send was armed when you left the chair.
- 4 Loops, retries, and runaway tasks Retries need a new tactic. Cap loops, watch the first run, find Stop first, and leave send and delete unarmed until a person approves.
- 5 Checking agent work without being an engineer Check the files, not the agent transcript. Verify the number, source, owner, and whether send already happened. Pass or fail in sixty seconds.