---
title: 'In the News: September 19, 2026'
description: "Maurice Heumann measured what his harness subagents were duplicating, turned them off, and found the reviewer's model mattered more than the worker's."
canonical_url: 'https://darkfactory.dev/news/2026-09-19-morning'
markdown_url: 'https://darkfactory.dev/news/2026-09-19-morning.md'
collection: news
date_published: '2026-09-19T07:45:00-04:00'
date_modified: '2026-09-19T07:45:00-04:00'
---

# In the News: September 19, 2026


Maurice Heumann ran four coding agents continuously for a month on a game decompilation project, then went back through about two gigabytes of session logs to find where the time and the tokens went. He turned subagents off, moved his context compaction threshold, and concluded that the model reviewing the code matters more than the model writing it.

## 1. Heumann audits his four-agent harness and turns the subagents off

**[235 Billion Tokens Later: Auditing What The AI Agents Really Did](https://momo5502.com/posts/2026-08-22-mw2-decompilation-audit/)** · Maurice Heumann · Personal blog, August 22, 2026

Heumann drew the measurements from session logs for the month-long run he described in August. Across 47 subagent runs, he compared every read of a file at least 5kB in size. More than one subagent read 36% of the distinct large files. Weighted by volume, 54.7% of what the subagents read from large files was duplication, and one status file was read 28 times by 21 different subagents. He disabled subagents, after which average commits per day moved from 243.9 to 311.1, a proxy he calls "not necessarily accurate".

Sessions had been reaching 93% to 100% of a one-million-token context window before compacting, at a median of 933,088 tokens, so Heumann retuned the harness to compact at 60%. Sonnet 5 recorded a 9.1% build failure rate against Opus 5's 23.9%. Once bugs flagged were normalised per 100 commits, the two workers made roughly the same number of mistakes: "it's the reviewer's model that matters, not the worker's". He puts the run's list-price-equivalent API spend at $85,207 against a subscription bill of roughly 200 euros, with about 89% of token volume served from cache.

**Why it matters:** Subagents are normally added to keep research out of the parent agent's context, and these logs show the duplicated reads that can result. Any harness that records file access can run the same comparison. Where one agent reviews another's commits, Heumann's measurements support assigning the stronger model to the reviewer role.