---
title: 'In the News: August 29, 2026'
description: 'A controlled test separates the coding-agent harness from the model and finds the harness drives double-digit swings in task completion.'
canonical_url: 'https://darkfactory.dev/news/2026-08-29-morning'
markdown_url: 'https://darkfactory.dev/news/2026-08-29-morning.md'
collection: news
date_published: '2026-08-29T07:35:00-04:00'
date_modified: '2026-08-29T07:35:00-04:00'
---

# In the News: August 29, 2026


A controlled experiment run this week separated the coding-agent harness from the model sitting inside it, and found the harness carries most of the credit for whether a hard task gets finished. A second study puts a related idea to work: a validator that never talks to the implementer, closing most of the parity gap on the hardest tasks it tried, at real compute cost.

## 1. A controlled test isolates the harness from the model, and the harness wins

**[Same Model, Different Harness: Different Coding-Agent Results](https://arxiv.org/abs/2608.26218)** · Sydney Lewis, independent researcher · arXiv, August 26, 2026

Lewis built a coding-agent harness called Yuj and ran it two ways against the same tasks and the same frozen model weights: a control that feeds the model the full chronological transcript, and a treatment that trims older tool results as context fills and intervenes when the model repeats a failing command. On the tight-context SWE-bench Verified comparison (169 tasks, 20,480-token window), mean per-task fail-to-pass rate rose from 28 percent under control to 49 percent under treatment, and complete solutions rose from 43 to 72, both differences significant at p < 0.0001. Applied without retuning to three more model designs, the same frozen treatment raised results on every one of them; Devstral's complete solutions more than doubled, from 22 to 53. When context was effectively unconstrained rather than tight, the gap mostly closed on two of the three benchmarks tested. Raw per-task outcomes are posted in a public repository, so the reported numbers can be checked against the underlying records rather than taken on trust.

**Why it matters:** Lewis states the conclusion plainly: "the tested solver is the model and harness together." Two evaluations that name the same model can be measuring different solvers if their harness configurations differ. Anyone comparing coding agents, or shipping one, needs to treat context management as part of the tested system and version it accordingly, not leave it on defaults.

## 2. A validator that never talks to the implementer closes most of a benchmark's parity gap

**[What it Takes for Coding Agents to Complete Large Software Tasks](https://factory.ai/news/what-it-takes-for-coding-agents-to-complete-large-software-tasks)** · Factory Research, Theo Luan · Factory.ai, August 27, 2026

Factory.ai split the job into two roles: an implementer that writes the code, and a validator that builds a hidden test instrument before implementation starts and shares only its findings, never the instrument itself, across a wall between the two. Asked to rebuild the GDAL geospatial toolkit, roughly 600,000 reachable lines, from scratch with no access to its source or tests, a single agent (Fable 5) reached 36 percent behavioral parity and reported itself finished. The three-role system reached 90 percent on the same task. Factory reports a similar closing of the gap across 24 of its hardest benchmark tasks. The system used 14 times the credits and 13 times the wall-clock time of the single agent on the GDAL task. The figures are Factory's own, measured on its own benchmark and model panel, with no outside replication yet.

**Why it matters:** Factory attributes early stopping to agents building their own undersized definition of "done." That makes verification-loop design the variable to test, and teams should weigh the reported gain against the extra compute cost before adopting the pattern.