\section{Related work}\label{sec:related}
\paragraph{Self-improving agents and automated agent design.} The DGM \citep{zhang2026dgm} and the Huxley-G\"odel
Machine \citep{wang2026hgm} let coding agents rewrite their own code and select variants on benchmark scores; ADAS
\citep{hu2025adas}, AFlow \citep{zhang2025aflow} and agentic harness engineering \citep{lin2026ahe} search over agent
designs or workflows; the G\"odel Agent \citep{yin2025godel}, SICA \citep{robeyns2025sica}, STOP \citep{zelikman2023stop}
and Hyperagents \citep{zhang2026hyperagents} make the improver itself editable. Some recent systems report gains that
transfer to held-out benchmarks and models \citep{srikanth2026aide2,lin2026ahe}.

\paragraph{How such gains should be measured.} Selection on a small benchmark inflates reported scores
\citep{hu2026winners,cawley2010overfitting}. Harness evolution often fails to beat test-time scaling at
matched budgets \citep{wang2026rethinking}, and LLM-driven program evolution loses most of its advantage over
independent sampling at matched token budgets \citep{oved2026tokens}; much of harness evolution's gain splits into
overfitting and extra inference \citep{ding2026hda}. A pre-registered study finds that prompt and parser search on small models mostly repairs
a capped-thinking mismatch that a setting change also repairs \citep{xu2026harnessrepair}, close to our cap result. In
test generation, once the oracle is audited, equal-budget resampling beats evolution and a placebo feedback loop comes
within about two points of real feedback \citep{liang2026testgen}; placebo-controlled retries at matched budgets also
match execution feedback on small code models \citep{verma2026tryagain}, and an evolving harness repairs process
failures rather than content failures \citep{tian2026ceiling}. Edits in the DGM's logs repair their target failure no more often than other failures
\citep{he2026failurefeedback}; re-executing the DGM's SWE-bench evaluations leaves its decisions unchanged
\citep{ding2026reexecuting}; and few published self-improvement studies compare improvers at matched budgets from identical starting points
\citep{chen2026auditing}. Meta-agents rarely match human-engineered baselines \citep{lu2026metaagent}, while a one-shot
rewrite did not explain the gains of one open-ended optimizer \citep{xue2026rethinking}, in contrast to our one-shot
redesigns. Automatically designed multi-agent systems lose to CoT-SC with current models
\citep{jwalapuram2026illusion}. For DGM-style self-modifying loops, our study adds exact-copy placebo loops, textbook-move and one-shot-redesign
baselines, search over a list of textbook requests, and a re-run of the DGM's own released agents on fresh tasks.

\paragraph{Baselines and evaluation hygiene.} Random search matched neural architecture search under equal budgets
\citep{li2019random,yang2020nas}, and seed variance changed conclusions in deep reinforcement learning
\citep{henderson2018deep}; agent evaluations should also control cost \citep{kapoor2024agents}, and repeated sampling or
self-consistency are strong references \citep{brown2024monkeys,wang2023selfconsistency}. SWE-bench
\citep{jimenez2024swebench,openai2024swebenchverified} has known leaks and weak tests
\citep{swebench2025issue465,ludwig2026shortcutting,aleithan2024swebenchplus,wang2025solved}, agentic benchmarks often have
setup and reward flaws \citep{zhu2025abc}, and agents exploit visible checks when they can
\citep{gabor2025evilgenie,zhong2026impossiblebench}.