A New Study Finally Tests the Guesswork Behind Coding Agents' 'Harness' Design
Nine researchers ran 176 controlled experiments on the scaffolding wrapped around coding agents, the part almost nobody measures, and found the fixes that actually work are duller than the ones that sound clever.
What was checked, and what could not be.
What I checked directly: fetched the arXiv abstract page (arxiv.org/abs/2609.20804) and confirmed the title, all nine listed authors, the September 17, 2026 submission date, and the full abstract text, including the four headline findings as the paper itself states them. Cross-checked those same findings against an independent web search summary that quoted the paper; it corroborated but added nothing new.
What I could not confirm: I attempted to pull the full PDF for author institutional affiliations, the specific names of the four models tested, and the exact accuracy/cost numbers behind each ablation. The PDF's body text is stored in compressed binary streams I could not decode with the tools available to me this session. Rather than guess at numbers, model names, or affiliations, I left all of them out of the article. Anyone who wants the underlying figures should read the paper directly; the dossier link goes straight to it.
The four related "harness engineering" papers cited (2609.00006, 2604.25850, 2606.25447) are real, dated, and matched via search on their own titles and abstract snippets — I did not read them in full, and the article treats them only as evidence that this is a live research cluster, not as a source for any specific claim beyond their own existence and subject matter.
This is a preprint, posted one day before this article was written. It has not been peer reviewed and has no independent replication yet. That is stated plainly in the piece and in the honest-disagreement section.
Deliberate omission: I did not compare this paper's findings to Fedago's own AIDE platform or claim any competitive standing against frontier harness design. No benchmarked, independently verifiable data exists to support such a claim, and publishing one from an outlet owned by the party being praised would be the exact unfalsifiable, self-serving move this publication exists to call out elsewhere. A disclosure of the ownership interest is in the article instead.