---
title: 'In the News: October 27, 2026 (Morning)'
description: 'A first-person software factory write-up caps retries and uses scripts where no model is needed; a source study of eleven coding harnesses finds none use an agent framework.'
canonical_url: 'https://darkfactory.dev/news/2026-10-27-morning'
markdown_url: 'https://darkfactory.dev/news/2026-10-27-morning.md'
collection: news
date_published: '2026-10-27T07:30:00-04:00'
date_modified: '2026-10-27T07:30:00-04:00'
---

# In the News: October 27, 2026 (Morning)


Two items this morning: a developer-relations engineer's account of building a small software factory with a trace of every decision, and a source-code study that reports what eleven production coding harnesses are built from.

## 1. A software factory builder cut cost by capping retries and replacing one worker with a script

**[Welcome to Glass Town: building a software factory with a paper trail](https://aaif.io/blog/welcome-to-glass-town-building-a-software-factory-with-a-paper-trail)** · Rizel Scarlett · Agentic AI Foundation (AAIF), October 14, 2026

Scarlett built a factory with five workers: a planner, a builder, a tutorial writer, a reviewer and a previewer. It runs on a persistent remote machine (a Sprite) and, in her words, it is a "transparent, well-lit factory" set against Steve Yegge's dark one. She reports two problems. A factory is new territory for coding agents, so it "took me hours to get it right." And failed runs still burn tokens: the factory kept pausing, she kept rerunning it, and she had hit her Codex usage limit while an automatic top-up kept the spend going.

Her fixes are specific. Every coding role moved to a cheaper model (gpt-5.6-luna) at low reasoning. She turned off automatic retries and gave each worker five minutes, so a failure stops the factory and leaves the retry decision to her. The preview worker became a plain script because it needs no model. For traceability she used the open-source Entire CLI, which links a recorded session to each Git commit, and added an AGENTS.md instruction to commit and push after every single file change. She ran two projects through it in parallel and calls it a success. The post gives no cost figures and no measured comparison, so the savings are her account, not a result.

**Why it matters:** The cost controls here are the portable part: a retry cap, a per-worker time limit and a script in place of a model call each turn an open-ended spend into a bounded one. The per-file commit rule is a cheap way to make a session reconstructable after the fact.

## 2. A source study of eleven coding harnesses finds none use an agent framework or embedding retrieval

**[Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents](https://arxiv.org/abs/2609.00006)** · Paul Barbaste and three co-authors, Wavestone AI Lab · arXiv, July 15, 2026

The authors read the source of eleven coding harnesses, including Claude Code, Codex CLI, Gemini CLI, OpenHands, Aider, Hermes and OpenCode, plus Databricks' Omnigent as a contrast case. They describe an 83-page paper with 13 cross-cutting observations and 29 recurring design patterns, and do not benchmark or rank the systems. Across roughly four million lines of code, they report that no runtime imports a general-purpose agent framework such as LangChain, LangGraph or AutoGen, and none retrieves code with vector embeddings; the field uses hand-rolled async loops plus ripgrep, tree-sitter and glob. SKILL.md skills lead MCP in adoption, 9 of 11 systems against 8 of 11.

One observation concerns loops: Mini-SWE-Agent's minimal linear loop reports results in the same range as OpenHands' event-sourced engine, so loop sophistication does not predict benchmark performance. The paper also source-diffs eight systems across one quarter and reports that Codex adopted Claude Code's hook vocabulary verbatim. It closes with 18 design recommendations and a 90-line minimal harness. These are the authors' findings from their own reading; this edition read the abstract, introduction and the loop sections, not the whole paper, and nothing here has been independently reproduced.

**Why it matters:** If you are choosing a stack, the reported pattern is that production harnesses sit on plain loops and deterministic search, with extension through skills and MCP. That is evidence about what shipped systems do, not about what works best.