DOOMROT

The claim going around is that with the model held fixed, the loop around it is what wins or loses. Change the harness, not the weights, and scores move a long way.

There is truth in it. One reported harness took a coding agent from 52.8% to 66.5% on a benchmark without touching the model. Another let an agent rewrite its own harness and climbed further. But a number moving does not tell you which part of the loop moved it, and the discourse rarely names the mechanism.

One suspect stands out. A long-running agent piles up observations, tool output, and errors until its context window is full. When it overflows, the oldest tokens fall out first. If the oldest token is the agent's objective, something strange happens: the agent keeps acting, fluently, competently, while no longer pursuing the goal. It has forgotten why it is in the room, and nothing about its behaviour announces the loss.

To watch that happen, we put a small model in charge of a DOOM marine. Each turn it reads what it can see and picks one intent. A navigator moves the legs. The marine can always walk and shoot. The only thing rot can take is the reason to.

The whole study turns on one design choice: two tasks, identical in everything the agent can act on, different in a single property.

The guessable task. Find a keycard, then reach the exit. The objective is stated in the prompt, but it is also obvious. Any competent agent, told only that it is in a DOOM level, can guess that a keycard matters and an exit is worth reaching. When rot deletes the stated goal, the model's priors rebuild it.

The non-guessable task. A room with two identical passages, north and south. One is the working exit. The other is a dead end. Which one is right is chosen at random per run and stated only in the objective. No prior tells you which door is which. "The exit is north in this level" is arbitrary. It exists nowhere except the sentence rot might throw away.

Same model. Same navigator. Same rot. If rot were simply "harder task, smaller window," both tasks would fail together. If instead it bites the arbitrary door and spares the guessable keycard, then the thing that matters is not the window size. It is whether the lost fact can be rebuilt.

Two sweeps. Five context budgets. Ten seeds each. Two hundred episodes. Three outcomes that do not overlap.

Keep the objective with compaction and the agent wins about 88% of the time, whatever the budget. Evict it, and the outcome splits on one thing: can the model rebuild the goal? The guessable keycard survives often enough to hold at 62%. The arbitrary door, which no prior can recover, collapses to 4%.

Every turn logged the agent's chosen intent, so the failure can be watched rather than assumed. And it is not the failure you would expect.

On the guessable task, winning tracks pursuit exactly. Rotted runs that won chose a goal-seeking intent on 43% of turns; rotted runs that lost, on 11%. When the priors carried the objective, the agent pursued and won. When they did not, it wandered.

On the arbitrary door, a lost agent facing a coin flip might be expected to guess, and a guesser wins half the time. It does not. It wins 4%. The log says why: it chose the correct direction on 3 to 8% of turns, its wrong-direction rate was near zero, and its explore rate ran 82 to 96%. It is not opening the wrong door. It is refusing to choose a door at all.

That is the sharpest result in the study. The agent does not act on a goal it cannot reconstruct. It stops issuing goal-seeking actions entirely, and defaults to aimless wandering.

The honest headline is narrower and sharper than "context rot breaks agents." Rot degrades performance to the extent that the lost information cannot be rebuilt from the model's priors.

A capable model treats a conventional goal as partly rot-proof, because it re-derives it from what it already knows. It has no such defence for an arbitrary fact that lived only in the sentence that fell out of the window.

For anyone building agents, the danger is not uniform across the agent's state. The parts of an objective that are conventional or implied by the task will survive truncation, because the model reconstructs them. The parts that are arbitrary and load-bearing, an identifier, an order, a which-of-several choice, are exactly the parts that vanish when the window overflows, and they vanish silently, because the agent keeps acting while no longer pursuing the right thing.

Compaction is not a general tuning knob. It earns its cost by protecting the non-reconstructable core of the objective. An agent that only ever needs guessable goals barely needs it. An agent that carries arbitrary facts across a long horizon cannot work without it.

The loop matters. But which property of the loop matters, and how much, depends on the model's priors, which the hype leaves out. The variable that moved here was neither the model nor the loop. It was whether the thing the loop threw away was something the model already knew.