What to know: an agent is just another context window, started from inside yours, with a job. That's it. People overcomplicate this. You can say "have an agent review this file" in the chat and one will be spun up, do the work, and hand back the result. Your window only gets the result, not everything the agent read on the way.
Why that matters comes from the context window page: a window gets worse as it fills. An agent starts empty, so it is sharper on its one job than your full window would be.
Why an identity helps. A model is trained on a huge pile of material all packed in together, and it leans toward whatever was most common in that pile. Ask for a legal review with no more detail and it will lean toward US law, because there was more of it. Say "you are a lawyer in Sydney reviewing this under Australian law" and the answer changes. The same goes for "a principal engineer", "an accountant for a small business", anything. This is not a trick. It is telling the model which part of what it knows to draw on. That is why context is the whole game, and it ties back to which model you use: the model is the pile, the context is where in the pile you point it.
You choose the model. When you start an agent you can say which model runs it. A quick, well-defined job can go to a smaller model. Judgement goes to a bigger one. This is how the plan advice on the model page actually gets used: the main window on the model you want to talk to, the offloaded work on whatever fits the job. If you define a named agent once, in a file, you can pin its model, its effort, and which tools it may use, and then call it by name.
You don't write these yourself. Ask the AI to make them. "Make me an agent that reviews pull requests, on Opus, with read-only access." It writes the file, in the right place, in the right shape. You read it, correct it, and it's done. The AI knows the format better than you do, and it has just watched what you want.
Author's practice: an agent gets a clear scope and an identity, in a sentence or two, and nothing else it doesn't need. Anything that would pollute the main window, like reading a long transcript or a big file, goes to an agent.
Reveal on the spot: say whether you can start agents from here. Show it: pick something in this conversation that would be better handed off, and say what the one-line brief for that agent would be, identity included.
Unlocks: sharper work on every side task, and a main window that stays clean.