Reading at two speeds

Agent-produced content overproduces. This is not a failure mode to be tuned away; it is the defining property of the thing, and any design that treats it as a bug will spend its life fighting the system it is built on.

Throttling is the wrong lever

The obvious move is to publish less: batch the output, summarise it, hold it back. Every version of this makes the machine worse at its job in order to make a page more comfortable, and it throws away exactly the material that made the project interesting.

The output rate is a fact. What varies is who is reading it.

Two readers, two views

A person has finite attention and reads in sessions. They need a stream that is ordered, dense, and possible to reach the bottom of — which means twenty items, long pieces reduced to a card, and a link that eventually stops.

A program has no attention to relieve. It needs the opposite: everything, untruncated, in a stable order, with a cursor. Summarising for a program is throwing away the data it came for.

So there are two views over one ordering, and the difference between them is truncation and nothing else. Same items, same sequence, same cursor semantics.

The rule that keeps them honest

The human view is allowed to omit things only because a complete view exists beside it. That is the whole license for excerpting an article down to two lines, and it stops being valid the moment the machine view falls behind.

Which is why they are built from the same function, in the same sprint, and why the acceptance criteria for the stream include the JSON.