---
title: 'In the News: October 24, 2026 (Midday)'
description: 'In a 320-run test, coding agents followed a stale CLAUDE.md or AGENTS.md over the code 90.1% of the time; one added sentence raised code-following from 4.2% to 91.4%.'
canonical_url: 'https://darkfactory.dev/news/2026-10-24-midday'
markdown_url: 'https://darkfactory.dev/news/2026-10-24-midday.md'
collection: news
date_published: '2026-10-24T13:40:00-04:00'
date_modified: '2026-10-24T13:40:00-04:00'
---

# In the News: October 24, 2026 (Midday)


A controlled study finds that coding agents obey an out-of-date instruction file even when they have noticed it contradicts the code, and that one sentence in the user's own file largely reverses that. Anthropic's Applied AI team also published guidance arguing that many enterprise agents should replace subagents with skills and tools.

## 1. Agents follow a stale CLAUDE.md or AGENTS.md over the code, and one sentence reverses it

**[Stale Instruction Files Override the Code in Coding Agents: A One-Sentence Fix](https://arxiv.org/abs/2610.30571)** · Xiaonan Xu (Georgia Institute of Technology) and Wenjing Wu (University of Colorado Boulder) · arXiv, October 22, 2026 (preprint, submitted to SN Computer Science)

The authors built 40 small Python projects in which the instruction file asks for one convention and the existing code uses another, such as print instead of logging. Claude Code (with Claude Sonnet 5.5) and Codex CLI (with GPT-6 Sol) each added a feature to every project, with and without a sentence in the user-level instruction file, for 320 runs. Without the sentence, 90.1% of the agents' choices followed the stale file and 4.2% followed the code. Claude Code named the conflict in its final reply in 79 of 80 runs and still followed the file. Codex CLI rewrote existing code to match the file in 20 of 80 runs, and the project's original tests then failed in 17 of those.

The added sentence, placed in the user-level file, tells the agent to follow the existing code when documentation and code disagree and to list each inconsistency it finds in its final reply. With it, 91.4% of choices followed the code, no Codex CLI run rewrote existing code, and tests passed in all 160 runs, up from 89.4% without it. The authors attribute the behavior to the file reaching the model as an instruction and the code reaching it as tool output, which models are trained to rank lower. The projects, run records and analysis scripts are listed as available in a public repository; we have not rerun them.

**Why it matters:** If these results carry over to real repositories, an old line in CLAUDE.md or AGENTS.md can cost a team more than a missing one, and agents will not warn you before acting on it. The study covers only synthetic Python projects, two agents and single unattended feature tasks, so test the sentence on your own repository before relying on it.

## 2. Anthropic's Applied AI team says many enterprise agents should fold subagents into one loop

**[Building agents: When to fold subagents into your main agent loop](https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop)** · Diane Sarkis · Anthropic, October 22, 2026

Anthropic says that across several enterprise deployments it compared, "a single agent with skills frequently outperformed both the one-prompt-for-everything design and the subagent design on quality," often at lower cost and latency. It lists the failure modes it sees in large organizations: subagent sprawl, a system prompt every team adds to, and evals tuned to each team's own subtask. It still recommends subagents for parallel work, for fresh context windows, and when a single loop degrades from context bloat or tool confusion.

The article gives a migration order: build a small end-to-end eval first, then fold in one subagent at a time and keep it only if the folded version scores worse. It also warns that folding one subagent into another lowered latency but broke a process that had been passing. The evidence is the vendor's own and is not quantified: it cites one unnamed customer that saw "a roughly 2x faster runtime" with a single loop, and says it cut Claude Code's system prompt by over 80%, with no data in the article.

**Why it matters:** The practical part is the procedure, not the headline claim: fold one subagent at a time against a fixed eval, run each task more than once, and compare which tasks flipped rather than the average score. Treat the performance claims as Anthropic's position until someone publishes numbers.