But if you need to access blocking I/O, do not do it in a coroutine.

Why not? And how else would you do it? Have two files being read in two coroutines separately? That's bad?

You can wrap blocking code to coroutines with own thread pool. After that you can use it easily with coroutines. And database io is good example where this technique will work

you could do it with builders as well using the DbPool from above

fun sqlQuery(query: String): Deferred<DbResult> {
    return async(BdPool) {  blockingSqlQuery(query)  }
}

I prefer suspend function in this case (easier to use). But version with

async

also useful in some cases, for example you can run sqlQuery in parallel, but practically I think it’s not very common thing and better to wrap to

async

in client code

Totally, if building a library for blocking like network or DB calls I would avoid taking a stance on how it is consumed at all, let the client decide.