Minimal harness matches elaborate ML engineering agent harnesses

What the research found

An arXiv paper (2609.40303, submitted 30 Sep 2026) finds that under equal time budget and the same frontier backbone, elaborate open-source MLE agent harnesses give no advantage over a single minimal-harness coding agent session (read, write, bash).

Claims and how well the evidence supports them

1. Under an equal time budget and the same frontier LLM backbone, open-source SOTA MLE harnesses provide no advantage over a single minimal-harness coding agent session 98%
2. The paper is arXiv 2609.40303, first submitted 30 Sep 2026 98%
3. Machinery layers become redundant in the coding agent setting per large-scale ablation studies 97%

Evidence

• arxiv.org https://arxiv.org/abs/2609.40303

Actionability rated low.