Agents use a separate context and won't pollute the main context.

So if you have a code review agent or a tdd agent checking the current commit if it matches some specs you have, they'll start a separate "subprocess" with its own context and return whatever they find to the main Claude context.