A coding agent is a model plus a harness: the code that decides what goes into the prompt, which tools exist, and when to stop. Hold the model, the tasks and the clock fixed, and change only the harness — keep the whole transcript on disk but shorten older tool results to their opening and closing lines once the prompt passes half the window. On 169 tasks with a 20,480-token window, the mean share of a task's failing tests that the patch fixes goes from 28% to 49%, and outright solutions from 43 to 72. Widen the window to 262,144 tokens and the two arms tie. The benchmark number you quote describes a pair, not a model.
Both arms store the same execution record. They differ in what the model is shown. The control serves the full conversation in time order and stops when the context fills, patch finished or not — the default in most agent loops. The treatment keeps that record and shortens the view: the newest four tool results stay whole, and older ones are cut to their beginning and end under a cap that halves as their age doubles. Nothing is deleted from the record; the trimming only starts once the estimated prompt reaches half the configured window. Drag the number of tool results and watch the two views diverge.
The treatment carries two more parts worth naming, because they are cheap to copy. A detector reads the execution record for two patterns — the same command failing again, and reading files without editing anything — and responds when it sees either. And a thin command layer rewrites test invocations, blocks forbidden commands, and stops an oversized setup step from flooding the window in the first place.
Three benchmarks, each run in a paired campaign: identical tasks, identical model weights, identical time limit, one changed harness. The score is the mean per-task share of a task's failing tests that the produced patch turns green — a partial-credit measure, which is why it moves before the all-or-nothing solve count does. Step through the benchmarks.
The harder benchmark shows the bigger multiple: on a 316-task set the solve count goes from 31 to 72. The feature-writing benchmark barely moves in complete solutions — 2 to 3 — while its partial-credit score nearly doubles, which is the signature of a task set where finishing is rare and progress is still real.
The obvious objection is that the treatment was tuned until one particular model liked it. It was not retuned: the same frozen settings were pointed at three further open-weight models with different internals — a dense transformer, a state-space hybrid, a dense linear-attention design — on the same 169 tasks and the same tight window. Every one improves. Step through them.
The spread is instructive: the weakest response, on the state-space hybrid, is still +6 points and a 1.6× lift in solutions. If a harness change of this size is available on four architectures without tuning, the harness is not a detail of the setup — it is part of the thing being measured.
This is the result that makes the rest credible, because it is the one an author hoping for a headline would rather not report. Run the identical paired comparison on the same 169 tasks with a 262,144-token window, and the two arms land on top of each other. The treatment is not making the model smarter. It is buying back room that the transcript was wasting, and when room is free there is nothing to buy.
One thing does survive into the wide regime: the treatment serves 7.2% fewer prompt tokens per turn for the same outcome. That is a cost line rather than a capability line — worth having, not worth restructuring your agent for.
The transfer is a sizing question. Take the window your agent actually runs with, the average size of a tool result in your loop, and the number of tool calls a typical task takes. If the full transcript would blow the window before the task ends, you are in the regime the paper measured, and trimming old tool output is the cheapest lever on the board. If it fits comfortably, the paper says you would be doing work for a 7% token saving.
Read the verdict line, not the meters. The paper's own claim is narrow and testable: this treatment is worth roughly twenty points of partial credit under pressure and nothing at all without it. The sizing above only tells you which of those two situations you are in.