---
title: 'In the News: November 15, 2026 (Extra)'
description: "Steve Yegge says he burned down five software factories and tells most teams not to build one; plus OpenAI's controls for persistent agents."
canonical_url: 'https://darkfactory.dev/news/2026-11-15-extra'
markdown_url: 'https://darkfactory.dev/news/2026-11-15-extra.md'
collection: news
date_published: '2026-11-15T10:10:00-05:00'
date_modified: '2026-11-15T10:10:00-05:00'
---

# In the News: November 15, 2026 (Extra)


Steve Yegge, who runs a software factory behind his own online game, now tells most people they don't need one, and says he burned down his previous five because rewriting them was cheaper than fixing them. The other two items are about containing agents that run unattended: OpenAI's published controls for its internal persistent agents, and a paper measuring which agent actions runtime controls actually see.

## 1. Yegge: factories need constant upkeep, and most teams don't need one

**[Steve Yegge on software factories, after two conference talks](https://x.com/Steve_Yegge/status/2121671447810888185)** · Steve Yegge, independent · X, November 14, 2026

In a long post distilling what audiences at two conferences found surprising in his talks, Yegge gives three positions. First, pace agent output to human speed. He says that even with about a dozen people working with his factory and a few hundred active players, his agents have "overwhelmed my players with features, my accountant with emails, my merge queue with tickets, and my machines with load," and "launched broken OSS without permission." Second, most teams do not need a factory: "Factories take an extraordinary amount of upkeep. They turn into garbage mountains unless you put a frightening amount of time and tokens into curating them. I had to burn my previous five factories down since it eventually became cheaper to rewrite them than to fix them." He argues for building over buying, because his factory has become "highly bespoke" to the game it builds and supports: "It's a company in a box." Third, he predicts that next year people will bring personal "entourages" of helper agents to work.

These are one operator's reported experiences, with no figures beyond his own counts. The post itself is not a talk transcript; the talks it summarizes have not been reviewed here. At about 10:00 EST on November 15, roughly 21 hours after posting, it showed 175.4K views, 1.7K likes and 2K bookmarks.

**Why it matters:** This is an operator describing a factory as a standing maintenance cost, paid in time, tokens and human leadership, not a one-time build. Anyone deciding whether to build one should budget for that upkeep. The concrete failures he lists, a flooded merge queue, overloaded machines and an unapproved release, are the places to put rate limits and approval gates first.

## 2. OpenAI publishes the controls it uses for its own persistent agents

**[Securing persistent agents like OpenClaw at OpenAI](https://developers.openai.com/blog/securing-persistent-agents-like-openclaw-at-openai)** · Kevin S Lin, OpenAI · OpenAI developers blog, November 12, 2026

OpenAI says it runs persistent agents across the company, including "hundreds of OpenClaw agents," and "over 500 agents that run proactively in the background." The post describes its setup. The OpenClaw gateway, which configures the agent, runs in a separate cluster from the agent harness, so only the agent workload is treated as untrusted. Agents run in the Codex sandbox with an organization-managed network allowlist, writes limited to their working directories, and secrets held by a credential broker the agent cannot read. A second model auto-reviews every agent action, which the post presents as a way to save human approvals for high-value decisions and avoid review fatigue. Each agent has its own identity tied to a service account, and every control-plane change and agent turn is logged. The post includes the `requirements.toml` and `config.toml` settings it uses, including removing the `danger-full-access` sandbox mode and denying cloud metadata endpoints.

**Why it matters:** This is a first-party control set with config you can copy, not general advice. The structural choice to note is the split: the component that configures the agent runs apart from the agent, which is what lets OpenAI treat everything the agent writes, including its own code, as untrusted.

## 3. A runtime control only sees the actions that pass through it

**[guardscope: Measuring Which Actions Runtime Controls Check in LLM Agent Systems](https://arxiv.org/abs/2611.15223)** · Kentaroh Toyoda, Alex Leung (Vulcan Research, AIFT), Bhuvaneswari U (Standard Chartered) · arXiv, November 2026

The authors list 15 paths an agent's action can take, from tool calls through a gateway to spawned background processes and direct network connections, and eight places a control can sit, from a gateway in front of the harness to the kernel. They then ran real controls, including the enforcement module of Microsoft's agent governance toolkit, agentgateway and the kernel-level ActPlane, with eBPF kernel tracing as ground truth. In their runs with a gateway in front of the harness plus the harness's own permission checks, both tool-call types were checked, but the eight direct-network, process-spawn, file-write and credential-access actions "take effect without any policy evaluation in every run." By their analysis, five of the 15 paths get past every placement short of the kernel, so stacking those controls adds detail about what is inspected but not coverage. Their workload exercised only 4 of the 15 paths; the rest rest on design analysis. They also measure coverage, not whether a policy decides correctly. The framework is released on GitHub.

**Why it matters:** If your agent controls live in a gateway or in the harness's permission prompts, actions the agent takes by shelling out or opening its own connections are not being checked. In the paper's runs, only a deployment that forced network traffic out through a controlled path stopped direct network calls, which is the job an organization-managed allowlist like OpenAI's is doing.
