---
title: 'In the News: October 29, 2026 (Extra)'
description: "A study of 1,886 coding-agent releases finds tool definitions make up 77.4% of Claude Code's first request and that fewer than half of prompt changes explain themselves."
canonical_url: 'https://darkfactory.dev/news/2026-10-29-extra'
markdown_url: 'https://darkfactory.dev/news/2026-10-29-extra.md'
collection: news
date_published: '2026-10-29T10:10:00-04:00'
date_modified: '2026-10-29T10:10:00-04:00'
---

# In the News: October 29, 2026 (Extra)


One item cleared the bar this cycle: a preprint that captured what five coding agents send their models, release by release, and found most of it is tool descriptions rather than the system prompt.

## 1. Researchers captured the first model request from 1,886 coding-agent releases; most of the text is tool definitions

**[Hidden in Plain Sight: an Empirical Study about The Instruction Stacks of Coding Agents](https://arxiv.org/abs/2610.37332)** · Felipe de Sant'Anna Paixão, Dhyego Tavares, Eduardo Almeida (Federal University of Bahia) and Iftekhar Ahmed (UC Irvine) · arXiv preprint, October 2026

The authors ran 1,886 npm releases of Claude Code, Codex, Gemini CLI, Qwen Code and OpenCode in sandboxes, sent each the same prompt, and recorded the first request each agent made to its main model. They call everything the agent adds around the user's prompt the "instruction stack". In the latest releases, tool definitions take 77.4% of Claude Code's request, 72.3% of Qwen Code's and 63.2% of OpenCode's. In one Claude Code capture, the 20 tool definitions run 57,346 of 74,078 characters, while the system prompt is 9,517. The authors' 21-character test prompt is between 0.02% and 0.07% of the latest requests.

Across releases, the stacks change in bursts. Of 1,867 measurable release-to-release transitions, 544 (29%) changed a segment or tool after the authors masked dates and version stamps. Claude Code 2.1.69 swapped 18 tool definitions for a single ToolSearch tool and cut its request from 77,601 to 20,380 characters; release 2.1.70 restored the definitions a day later. The authors also found a Claude Code directive to append a Co-Authored-By trailer to commit messages in 492 of 493 captures, and a comment policy that moved from "DO NOT ADD ***ANY*** COMMENTS unless asked" to "match its comment density" starting in release 2.1.154.

For the four open-source agents, the authors traced 496 change units to commits and pull requests. They report that 90% state some reason, but only 236 (48%) explain the change to the prompt stack itself. Safety was cited in 13 units. These are stated motivations coded with an LLM and checked against human labels on a sample, not verified intent, as the authors note.

The study covers only the first request, in an empty working directory, on default settings. The authors say it does not establish effects on correctness or cost, and that their replication package link is anonymized in the preprint; we did not open it.

**Why it matters:** If you compare agents or rerun a published agent result, the model name and your prompt do not describe what the model read. Record the exact agent release, because the tool descriptions can change the input by tens of thousands of characters between versions. If you mine repositories for agent-written commits, a trailer that one release asks for and another drops changes what a marker search finds. The paper measures what changed, not whether it changed outcomes.