---
title: 'In the News: November 18, 2026 (Midday)'
description: 'Kurtis Welch details a one-person agent factory that shipped 30 libraries, with outside release gates and 16 of 16 releases approved by tap.'
canonical_url: 'https://darkfactory.dev/news/2026-11-18-midday'
markdown_url: 'https://darkfactory.dev/news/2026-11-18-midday.md'
collection: news
date_published: '2026-11-18T13:05:00-05:00'
date_modified: '2026-11-18T13:05:00-05:00'
---

# In the News: November 18, 2026 (Midday)


Kurtis Welch says the agent factory he runs by himself has shipped 30 open-source libraries since early October, including a SQLite-compatible database engine and a deterministic physics engine. The useful part of his write-up is the operating detail: who may merge, which gates no agent can relax, and how untrusted GitHub issues reach the engineers. Every figure in it is his own.

## 1. A one-person software factory publishes its output and its rules

**[What my software factory built: 30 libraries, from a database engine to a game engine](https://recursiveai.org/blog/what-my-software-factory-built/)** · Kurtis Welch, Recursive AI · recursiveai.org, November 18, 2026

Welch reports 30 C# libraries on NuGet. 28 reached 1.0 between early October and mid-November, with nearly 1.9 million lines of library code and 241 releases, 183 of them on libraries already past 1.0. The set includes OneLite, a SQLite 3.53.4-compatible engine that he says passes all but one of nearly six million sqllogictest records, and OnePhysics, which he says produces identical world hashes in 165 of 165 scenarios across Linux, macOS, Windows and the browser. The source is public in the recursive-ai-oss GitHub organization, which holds 35 repositories created between October 6 and November 11.

The structure is explicit. A planner agent designs each library through multiple-choice questions, and a project manager agent answers them unless only Welch can. A build lead splits the work among engineer agents, each in its own git worktree. Engineers "can't push, tag or publish"; the lead checks their work and merges it. Runs start with no memory, so state lives in files: each library's PLAN.md is its lead's memory, and the project manager role has passed through 12 agent instances since November 4 by handoff note. Release gates come from outside the factory where possible: a standard's own conformance suite, a reference tool run as a black box, or real applications opening the output. In Welch's words, "No agent can relax a gate."

His own role, by his count for November 4 to 17: 72 board items reached him, he answered 55, 47 of them with a single tap, and he approved 16 of 16 release cards by tap. One case shows the loop end to end. Three GitHub issues from outside users on OnePdf, filed November 10 and November 16 and now closed, became release 1.23.0. The library's lead read the issue text once, "as data," and wrote briefs, so the engineer agents never saw it. Welch writes that he "wrote no code, reviewed no diff and wrote none of the replies."

We have not run the conformance suites or benchmarks behind these figures. On Hacker News the post had 14 points and 37 comments at 13:03 EST today, and commenters asked how the libraries compare with the originals they reimplement. One put it plainly: "The factory is only as valuable as its products."

**Why it matters:** Welch describes a factory he runs without reading diffs, and the controls are specific: merge rights sit with a lead agent, release gates are set outside the factory and no agent can loosen them, untrusted issue text is turned into briefs before engineers see it, and the human mostly approves releases. The quality claims still need outside testing. The controls can be copied and checked in your own setup now.
