Teams on a Fixed Model: Two Months of Harness Experiments with The One
How much can an agent-team harness add when the model is fixed? We report eight weeks (28 September to 20 November 2026) of controlled experiments on The One, a product that runs teams of language-model agents on a person's computer. Three programs ran back to back: V1 on a task suite of our own (over 120 batches), V8 on Terminal-Bench 4.0 through the Harbor framework (315 jobs and 6,223 graded trials, with 30 tasks held out and never read), and a final Terminal-Bench 4.1 run under the Artificial Analysis protocol. The configuration we kept, a single lead agent on GPT-6.2 Sol at maximum reasoning effort, scored 60.9% on the 63 tasks we could run and 58.1% with the three GPU tasks counted as failures. That ties the model vendor's own agent (Codex) on the same model and effort while using about 30% fewer input tokens, at $1.35 per attempt at list price. Three findings carry the paper. Append-only context cut cost by a third to a half with no change in accuracy. On tasks that one context can hold, a single strong owner matched or beat every team structure we built, and same-model second opinions confirmed the shared misreading instead of catching it. The model's choice at a decision point moved when facts were placed in front of it, never with advice. Most levers that won on the tasks they were designed on failed on held-out tasks; we describe the held-out discipline that caught them, and report the negative results in full.
Authors
- Ziyang Zhang
Institutions
- Zero Emissions Resource Organisation (NO)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24410339
- Primary Topic
- Artificial Intelligence Applications
- Type
- preprint