---
title: 'In the News: October 16, 2026 (Midday)'
description: 'A study of 600 coding-agent runs finds turn count, not tool-output size, drives cost; a new benchmark scores 19 models on building harnesses.'
canonical_url: 'https://darkfactory.dev/news/2026-10-16-midday'
markdown_url: 'https://darkfactory.dev/news/2026-10-16-midday.md'
collection: news
date_published: '2026-10-16T13:25:00-04:00'
date_modified: '2026-10-16T13:25:00-04:00'
---

# In the News: October 16, 2026 (Midday)


Two preprints this cycle, both about the loop around the model. One says the usual cost-cutting targets, long tool output and expensive models on every turn, are not where agent runs spend their money. The other scores 19 models on building a coding-agent harness from a specification, and the best one fails nearly half the tests.

## 1. A study of 600 coding-agent runs finds turn count explains 84% of cost variance

**[Turns, Not Observations: Cache-Aware Loop Engineering for Coding Agents](https://arxiv.org/abs/2610.15093)** · Pramod Kumar and Raj Kumar Pandey · arXiv, October 10, 2026; published in the proceedings of the PromptOps and Vibe Coding workshop at ASE 2026

The authors re-scored 600 publicly released coding-agent trajectories from two agent scaffolds and two open-weight models. Median tool observations were small, 118 to 286 tokens, and the large build logs often used to justify output compression sat near the 99th percentile. Prefix caching served 98.6% of input tokens, so pricing every input token at the flat rate overstated loop cost by roughly 7x on average (7.1x in the conclusion).

The cost driver was the number of turns. Turn count explained 84% of the variance in log trajectory cost; mean observation size explained 4%. The most expensive runs had the most turns and the smallest observations. The authors also report that model delegation, sending selected turns to a cheaper model, raised cost in half of the settings they tested, close to the break-even condition they derive. The paper releases a tool, TurnCost, for re-scoring trajectory sets under different caching and pricing assumptions.

The authors list their own limits. The corpus was built as distillation data and excludes runs that hit an iteration limit, so the observation-size medians are "indicative rather than exact." The turn-count relationship is partly mechanical, because cost is a sum over turns, though cost grew faster than turn count (exponent 1.23). The models are open-weight, not frontier proprietary agents. They also say they have not shown that turns can be removed without hurting task success. We read the abstract, introduction, limitations and conclusion; we did not check the per-setting results behind the delegation finding.

**Why it matters:** If you are tuning an agent loop for cost, this is evidence to count turns and cache behavior before you spend effort shrinking tool output or routing steps to cheaper models. The open question is the one the authors name: which turns are removable without losing the result.

## 2. Doc2Harness scores 19 models on building a coding-agent harness; the best passes 53%

**[Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents](https://arxiv.org/abs/2610.20344)** · Yuyao Ge and co-authors, Institute of Computing Technology, Chinese Academy of Sciences, with Tsinghua, Peking and Southeast universities · arXiv, October 14, 2026

Each model is given a development document consolidated from the user documentation of twelve frontier harnesses, plus a skeleton that declares interfaces only, and has to build a complete harness. The frozen harness is then run end to end against 328 test cases, driven by a scripted mock model, with results read from requests sent, files on disk, exit code and command output. Positive controls are there to catch tests that pass because the feature never ran.

Across 19 models, each building three times, GPT-6 Astra scored highest at 53.05% and Hy3 lowest at 27.44%. Opus 5 scored 46.65%. The top two split by scenario type: Opus 5 led on ordinary scenarios, 80.68% to 72.46%, while GPT-6 Astra led on boundary scenarios, 48.97% to 37.02%. Permissions was the category that separated models most, with a 50.8-point gap between strongest and weakest.

The architecture finding is the part a practitioner can use. Controlling for code size, harnesses split into more modules scored lower (partial correlation of -0.423), and the authors say failures cluster around how cross-cutting functionality is divided. This is an association across 57 builds. It is the authors' own benchmark, still a preprint, and each build allowed only up to three submissions; the test variants derived from one seed are correlated, which the authors flag. We read the results, architecture and limitations sections but not every part of the test-construction method.

**Why it matters:** This is a measurement of how well current models build the kind of harness the factory depends on, and it shows that gap is large and uneven by category, permissions most of all. Treat it as a first reference point rather than a ranking, and note that the one design lesson it offers, keep cross-cutting behavior in fewer modules, is a correlation, not a tested fix.