Aggregation is the server's half of the round: taking the updates from all participating clients and combining them into a new global model. The simplest and most common way is a weighted average — each client's contribution is proportional to how much data it trained on.
Why weight by data size? A client with ten thousand examples has more evidence about the world than a client with ten. If both counted equally, the tiny client would pull the shared model just as hard as the large one. Weighting by example count makes the average approximate what you would get if all the data had been pooled — without ever pooling it.
Aggregation is also where things can go wrong. If clients' data is very different (the non-IID problem from Week 1), their updates point in different directions, and the average can land somewhere that pleases no one. And if a malicious client sends a poisoned update, naive averaging absorbs the poison — a problem the security track studies in depth.
Aggregation turns many local lessons into one shared model. Give it a name and a formula, and you have FedAvg — tomorrow's topic.