---
title: 'In the News: November 2, 2026 (Extra 2)'
description: 'Factory reports its harness, which writes programs to call tools, scored higher at lower cost on Terminal-Bench 4 in its own runs.'
canonical_url: 'https://darkfactory.dev/news/2026-11-02-extra-2'
markdown_url: 'https://darkfactory.dev/news/2026-11-02-extra-2.md'
collection: news
date_published: '2026-11-02T16:04:00-05:00'
date_modified: '2026-11-02T16:04:00-05:00'
---

# In the News: November 2, 2026 (Extra 2)


Factory published numbers on what happens when an agent writes a program to call its tools instead of calling them one at a time. A grey-literature review of 76 practitioner sources catalogs what teams say they do about understanding lost to AI-written code.

## 1. Factory says its harness, which runs tool calls as programs, scored higher at lower cost on Terminal-Bench 4

**[Droid 1.0: Building a Frontier Harness](https://factory.com/news/frontier-harness)** · Factory Research · Factory blog, October 29, 2026

Factory made programmatic tool calling the default in Droid 1.0. The model writes JavaScript that runs in a sandbox with no file, network or system access of its own, and one model request can trigger hundreds of tool calls. Raw tool output stays inside the program, and the model reads only what the program prints. In Factory's Terminal-Bench 4 runs (October 23 to 28, three trials per task, compared only on tasks both harnesses passed in all three trials), Droid with Claude Opus 5.5 scored 65.2% at $4.73 per passed task, against 57.6% for Claude Code. With GPT-6.1 Sol it scored 57.6% at $0.79, against 54.0% for Codex. Factory says script mode cuts Droid's cost by 15% on Opus 5.5.

The post's worked example is a Linear cleanup of 151 issues: 312 tool calls, 2 model requests and about 61,000 tokens. Factory estimates direct calls would have taken 36 requests and about 3.1M input tokens, modeled at 10 calls per request. It also reports that a typical production request reads 98% of its input from cache, across 20.4M Droid CLI requests in the seven days to October 22, after Droid began appending newly discovered tools to the end of the request instead of inserting them into the tool list.

Factory built the harness, chose the settings and ran the comparison, and we did not rerun it. Some trials were rerun after provider failures, and moderated Opus 5.5 trials were rerun on Opus 5. Devin tasks where it stopped to ask the user were counted as passed if Codex passed them. The post also says Droid takes 10% longer than Codex with GPT-6.1 Sol.

**Why it matters:** If you run your own agent loop, the claim to test is narrow: how much raw tool output you paste back into context, and whether your request prefix stays stable between turns. Both can be measured on your own sessions before you accept a vendor ranking.

## 2. A review of 76 practitioner sources catalogs how teams try to prevent and detect lost understanding of AI-written code

**[Cognitive Debt in the GenAI Era: Listening to Practitioners' Voices on the Loss of Understanding](https://arxiv.org/abs/2610.37888)** · Theocharis Tavantzis (Aalborg University) and eight co-authors, including Margaret-Anne Storey (University of Victoria) · arXiv preprint

The authors reviewed sources published from February 2025 to September 2026: 58 practitioner articles, 11 company or vendor articles, five arXiv preprints, one industry report and one technical document. No forum threads passed their quality filter. They define cognitive debt as the accumulated gap between the understanding a team has and the understanding it needs for current or future work. Their finding is that it can build up unnoticed behind working code and passing tests when developers accept AI-generated solutions without building understanding.

The preventive practices practitioners propose include comprehension gates, review-sized changes, externalizing intent and executable guardrails. Proposed signals include reliability metrics and cognitive-debt audits.

This is a synthesis of what people wrote, not a measurement. The authors say they do not estimate prevalence and did not compute inter-coder agreement, and they used Claude Fable 5.1 for part of the source search. We read the abstract, the method and the opening of the results, not the full results.

**Why it matters:** It is a list of what other teams report trying, which is useful if your agents merge changes faster than anyone reviews them. Because it counts what practitioners wrote and not what worked, treat each practice as a candidate to test.