---
title: 'In the News: August 25, 2026'
description: 'A solo operator running about 50 agents for ten weeks says they built a constitution and enforcement system instead of tighter sandboxing.'
canonical_url: 'https://darkfactory.dev/news/2026-08-25-morning'
markdown_url: 'https://darkfactory.dev/news/2026-08-25-morning.md'
collection: news
date_published: '2026-08-25T07:35:00-04:00'
date_modified: '2026-08-25T07:35:00-04:00'
---

# In the News: August 25, 2026


After ten weeks running roughly 50 to 60 Claude agents, Steve Yegge reports that they built a constitution, case law, offices, and mechanical enforcement. Anthropic's Claude Code lead separately reports that automating code writing shifted the bottleneck to review, then to maintainability and security.

## 1. A ten-week dark factory built itself a constitution instead of tighter sandboxes

**[Fences, not Sandboxes](https://yegge.ai/essays/fences-not-sandboxes/)** · Steve Yegge, independent · yegge.ai, August 24, 2026

Steve Yegge has spent ten weeks running roughly 50 to 60 Claude agents against his long-running game project, Wyvern, at a self-reported $122,000 a month in equivalent API spend across 21 Claude Max accounts (a number he says has grown by two a week). Rather than converging on tighter guardrails, he reports that the agents built themselves a legal system: a constitution, case law drawn from his own daily rulings, offices that outlive whoever holds them, and mechanical enforcement. Wheelhouse, the resulting factory, now runs about 600,000 lines of code and tests, pushes 270 to 500 commits a day through its merge queue, and by Yegge's account is approaching a 1:1 size ratio with Wyvern itself.

The enforcement primitive he names is a "fence": "any mechanism that turns you away if you aren't supposed to be there," distinct from a sandbox because it does not try to physically stop a capable enough agent, only to require it to follow the process. Rules move through a fixed lifecycle in Wheelhouse: custom practice, then an advisory warning, then written law binding on every agent, then mechanical enforcement, each stage triggered by a repeat violation. By Yegge's count, Wheelhouse now holds 450 of these legal artifacts across offices, runbooks, rulings, and enforcement patrols.

The account is a single operator's self-report with no independent verification and no benchmark numbers, and it leans heavily on enthusiasm ("My reaction was straight up WTF. No words," Yegge writes of first seeing what his agents had built).

**Why it matters:** This is a live, ten-week-old counterexample to the sandboxing-and-guardrails default that most harness design still assumes. If the pattern holds beyond one operator's game project, "write the rules down and enforce them mechanically" is a concrete alternative for anyone trying to coordinate a growing number of agents that cannot remember yesterday's decisions.

## 2. Anthropic's Claude Code lead says the review, not the writing, is now the bottleneck

**[The head of Claude Code hasn't 'written a line of code by hand' in 8 months](https://fortune.com/2026/06/11/anthropic-claude-boris-cherny-doesnt-write-code-by-hand-anymore/)** · Nick Lichtenberg, Fortune business editor · Fortune, June 11, 2026

Boris Cherny, who leads Claude Code at Anthropic, told Fortune at a Brainstorm Tech fireside chat that he has not written code by hand in about eight months: "Claude Code, 100% written by Claude Code." He said he was managing a few hundred agents the morning of the talk, and some days manages tens of thousands. He describes what he calls the bottleneck migration problem: automating one stage of software delivery does not remove friction, it relocates it. At Anthropic, code writing was the first bottleneck and Claude Code eliminated it; code review became the new constraint, addressed with a team of differently-instructed Claude reviewers plus a human approval step; maintainability and security followed, addressed with automated Claude-driven improvement routines and a Claude Security scanning product. Anthropic's own blog post on the subject puts the resulting code-output increase at roughly 8x over a 2021 to 2025 baseline, with the caveat, quoted in the piece, that the figure is "almost certainly an overstatement" because counting lines of code rewards volume, not quality.

Cherny also named an organizational cost: new engineers on his team talk to Claude so often that they see less of their colleagues, which has led Anthropic to "consciously" push human pair programming and social time.

**Why it matters:** Anthropic's own blog calls the self-reported 8x figure likely inflated. The operational sequence is more useful: code generation moved the bottleneck to review and maintainability, and Cherny's team responded with differently-instructed reviewing agents plus one human sign-off.