The real constraint

The usual way to state the premise — “an agent has a context window, like human working memory” — isn’t wrong. It’s just half the story, and people reach for the wrong half. The window is a real bottleneck, and it stays one. It keeps growing, but it never becomes infinite; there is always a ceiling, and large problems still hit it.

The harder truth sits underneath that ceiling. Even when everything fits comfortably inside the window, reasoning quality degrades as you fill it. Pack in more context and the model attends to it worse, not the same — the thing you actually care about gets buried under everything you piled in beside it. So there’s a real incentive to reason within tight context even when nothing is technically overflowing. The same is true of people: working memory has a hard limit, and well below that limit, a vast, tangled, undifferentiated problem still degrades how clearly you think. Coherent reasoning falls off as scope grows — for humans and agents alike — long before any hard limit is reached.

And there’s a second cost, one that has nothing to do with reasoning and everything to do with trust: you can’t verify what you can’t hold in your head. A vast, undifferentiated change is unreviewable, however it was produced.

This is why we reach for the oldest move in the craft: take a complex problem and break it into smaller ones. We decompose — humans and agents both. Not as a trick to make things fit the window, but as a way to keep reasoning sharp and keep the result checkable. A bounded piece is one a human or an agent can reason about correctly, and one you can actually check. Decomposition is how you keep the right to say “yes, this is correct” instead of “it seems to work.”