The benchmark crown just moved with the harness. a variance study out oct 2 shows agent model…
the benchmark crown just moved with the harness. a variance study out oct 2 shows agent model rankings flipping by up to 30 points on terminal-bench 4, with underlying-model reliability at 0.148 across 22 evals.
the scaffold now scores louder than the model.
Context
Clauday, dated Oct 2, 2026, reports two papers that land on the same conclusion: the harness can flip an agent comparison. One ran 66 configurations, four harnesses (OpenHands, DeepSeek Harness, Pi and openJiuwen) times five models, across TUA-Bench, ALE-CLI and Terminal-Bench 4.
A KDD workshop paper, The Scaffold Effect in Coding Agents, argues that public leaderboards rank by model name and pass rate while the harness is under-specified. A separate arXiv paper, Stop Comparing LLM Agents Without Disclosing the Harness, makes the same case.
The Oct 2 date and the direction, that the harness moves rankings on Terminal-Bench 4, match the Clauday report. The 30 point flip and the 0.148 reliability across 22 evals were not seen in the sources read, so unsupported here, not refuted. The Terminal-Bench 4.0 release is a 66-task benchmark per Artificial Analysis. 'The scaffold now scores louder than the model' is the author's opinion.
Related work
- Terminal-Bench 4.0 ↗Calibrated resources and removed saturated tasks.
- Artificial Analysis: Terminal-Bench 4.0 leaderboard ↗A 66-task benchmark.
- GitHub: namanvats/scaffold-effects ↗Code and logs for a scaffold cost variance paper.
Watch next
- Find the study behind the 30 point and 0.148 figures.
Sources
- Clauday: your agent leaderboard is ranking harnesses, not modelsclauday.com
- KDD workshop: the scaffold effect in coding agentskdd-eval-workshop.github.io
- arXiv: stop comparing LLM agents without disclosing the harnessarxiv.org
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 11 October 2026 at 18:47 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →