Some data arrives in pieces: pages from an API, lines from a stream. An async generator is an async def that uses yield. It can await between items, and the caller reads it with async for.

import asyncio

async def countdown(n):
    while n > 0:
        await asyncio.sleep(0.01)
        yield n
        n -= 1

async def main():
    async for n in countdown(3):
        print(n)
    evens = [n async for n in countdown(6)
             if n % 2 == 0]
    print(evens)  # [6, 4, 2]

asyncio.run(main())

Like an ordinary generator it is lazy: it runs only when the loop asks for the next item. If the loop stops early, the rest is never fetched.

A plain for cannot read it, and neither can list(). Use async for, or an async comprehension as above, inside a coroutine.

Write an async generator paginate(get_page) that awaits get_page(1), get_page(2) and so on, yields every item of every page, and stops at the first empty page. Write async def take(items, n) that returns a list of the first n items of an async iterable, without reading any further.

The explanation opens here once every check passes.