An autonomous machine-learning engineering paper dated September 30, 2026, and listed on Apple’s research site in October compares a minimal coding-agent setup with more elaborate systems. Its authors are Kirill Brilliantov, Alejandro Hernández-Cano and Emmanuel Abbé; the paper lists EPFL and Apple affiliations and notes work done during an Apple internship.
The baseline, called Malena, is a single coding-agent session with direct file access and code execution. The researchers held model backbone, hardware and time budget constant when comparing systems. Their stated result is that this minimal setup matched or outperformed the tested open-source systems on current ML engineering benchmarks.
The study includes MLE-bench and NatureBench. Its central point is methodological: comparing systems powered by different models can make a model improvement look like a system-design improvement. Matched backbones help separate those effects.
The authors found direct access to the execution environment to be the largest measured intervention. Additional search, autonomy and multi-agent coordination did not produce a statistically significant gain in the tested coding-agent setting. They also note weaker backbones can still benefit from structured workflows.
This is a research result with limits. The paper discusses expensive evaluations, modest seed counts and run-to-run noise. It does not prove that a single agent is better for every production task, collaboration pattern or safety-sensitive workflow. It also does not announce a new Apple consumer product.
For builders, the supported takeaway is to include a strong simple baseline and compare under matched budgets and models before attributing gains to added complexity. We have not reproduced the experiments.