What is the consumption problem?

AI makes output cheap. A person’s attention does not grow.

Play the animation Show the finished frame

The consumption problem: output grows, attention does not, so work is bound to its sources and checks Four panels: output piles up past one person's attention; items then link to sources; a changed source marks items stale. A person beside a fixed window of five cards and an equal pile of five. attention output Output is cheap to make. Attention stays the same. The pile of cards now stands far above the same fixed window; the cards above it are dimmed. attention output Output keeps growing. The window does not. A grid of cards, each linked by a line to a source tile and tagged checked. checked checked checked checked checked checked checked checked checked checked checked checked checked checked checked checked policy.md source source Each item links to its source and shows its status. One source tile is highlighted and the cards linked to it change from checked to stale. stale checked checked checked checked stale checked checked checked checked stale checked checked checked checked checked checked checked checked policy.md changed source source A source changes. Work that depends on it goes stale.
  1. Output is cheap to make. Attention stays the same.
  2. Output keeps growing. The window does not.
  3. Each item links to its source and shows its status.
  4. A source changes. Work that depends on it goes stale.
Illustration of the recorded example, not a live run.

Output outruns the reader

In my own projects, agents produced summaries and instructions faster than I could check which ones still held. Over six weeks in 2026, agents merged about 2,300 pull requests in one of my repositories, roughly 50 a day. The study.

The same gap shows up elsewhere, and it is older than AI. Congress gets more constituent messages than its staff can handle (Congressional Management Foundation, 2021). AI makes it cheap to fill feeds and websites with slop (Stanford and Georgetown researchers, 2024).

Bind work to its sources

More output needs a different kind of help: a link. In the Consumption Engine, recorded work is bound to its sources and checks. In the public example it shows as checked or stale.

When a source changes

When a check finds that a source I declared has changed, the work that depends on it is held back until it is checked again. The old work stays in history. The new work gets its own check.

The public example: a policy changes from three retries to two. The three-retry answer goes stale. The two-retry answer carries its own check. Watch it or run it in plain Python.

Who decides

A person still judges whether a source supports a claim. The engine keeps each result, its source and its check where that person can see them.

Read the research brief.