You'll finish with: You understand isolated context and when it helps.
You ask Claude to research a prospect. It runs twelve searches, reads nine pages, and dumps all of it into your conversation. Somewhere in there is the answer. Now every message after that is dragging the whole pile along behind it.
Same job, two ways
A worker with its own conversation, its own instructions, and its own limits. You hand it a job. It goes away, does the work in its own space, and comes back with the result.
Think of it like sending an apprentice to price up a job. You don't want to watch every phone call they make. You want the number when they're done.
Quick questions. A one-line answer doesn't need a delegate — you'll spend longer describing the job than doing it.
Build two or three, not all five. Start with the ones matching your loudest problems, then add the orchestrator that routes between them — that's chapter 6, the last one in this course. A pile of agents with nothing coordinating them is just a folder of files.
You can name one job worth delegating, and say what the finished output should look like in one sentence.
If the sentence comes out as "help me with clients" or "make things faster", that isn't a job yet — it's a wish, and an agent built on it will wander. A job you can delegate has a trigger, a finish, and an output you could describe to someone who doesn't work here.