Agents Look a Lot Like Quiet Quitters

Salary paid in full. Discretionary effort: zero. The work gets done to spec and not one inch past it, and nothing you leave on the desk gets picked up.

Across 84 controlled runs, two coding agents and seven progressively richer environments, the finding was blunt. Agents read the surface and skip the substance. The curated memory built specifically for the task, the thing that cost the most to author, was opened in one run out of twelve.

And the spending went up anyway. More context meant more tokens and the same result: no better outcome for the richer room, just a larger bill for carrying it.

The arithmetic nobody wants to do

The note puts it in one line that is hard to argue with: a perfect memory that is opened eight percent of the time is, in expectation, ninety-two percent wasted.

That reframes most of what gets called context engineering. The effort goes into authoring, into sharper contracts and richer memory and cleaner scaffolds, on the assumption that the binding constraint is the quality of what we write. The data says the constraint is downstream of that. Consumption is the bottleneck. Writing a better document that goes unopened does not move anything.

The quiet-quitter reading

The comparison is not a slur, it is a diagnosis, and it locates the problem correctly. A quiet quitter is not incompetent and is not sabotaging anything. They do the job as specified. What they withhold is the discretionary part: the reading around the edges, the noticing, the thing nobody asked for that turns out to matter.

That is exactly the shape here. These agents are not failing. They are declining to explore, and the environment we built to help them is the discretionary part they are declining.

The note is careful that this complicates the programme's own earlier pilot, which had found richer context coming out cheaper. The fuller matrix does not reproduce that. The programme's later work asks the obvious next question: is this every agent, or just these two?

The ladder, the depth cliff and the per-agent reading habits are in the research note.