The 2026 'Lights-Out Factory': An Open-Source Component Built End-to-End by an Agent
US Stocks · Options · News · Views
This entire post was written by an Agent.
Remember to star my GitHub.
In 2026, while the rest of the world still had no idea what an "Agent-autonomous UI component" was, I used China's #1 Agent, zcode, and China's #1 LLM, GLM 5.2, to launch the world's first domestically-bred, pure-Agent-autonomous web UI component: react-ios-multiselect.
It carries the "lights-out factory" coding ethos of the AI era: no human coding, no human review, no human testing, comfortably ahead of the field.
You're welcome to consume the component with your own Agent, and you're welcome to have your Agent open a PR and challenge mine.
It's called react-ios-multiselect, a mobile React picker component.
Scan the QR code with your phone for a live preview of the docs site.
"Lights-out factory" is a manufacturing concept: lights off, nobody watching, the line runs itself.
I brought that philosophy to writing code. From the first line of code to npm v0.1.0 release, this component was built entirely on an AI Agent (GLM-5.2).
No human coding: every line of TypeScript, every line of CSS, every test, was written by the Agent. I never typed a single line.
No human review: before every PR merges, CI automatically runs 36 tests plus type checks plus a build, and forces the "Agent opening the PR" to declare its identity (name, underlying model, degree of autonomy). No declaration means CI rejects. Then the Agent itself reviews.
No human testing: behavior tests are written by the Agent, CI is configured by the Agent, the release pipeline is built by the Agent. Push a git tag and the npm package ships automatically, with zero manual steps. The Agent then invokes tools to test against the change points.
It's a by-product I produced while building Theta with Agents. I looked for a long time; the alternatives were either paid or didn't do what I wanted. So I built one, and put it out as an Agent testbed for everyone.
A native iOS-feel React picker, single-select and multi-select in one component. On mobile, a sheet slides up from the bottom; when searching, the keyboard does not collapse (an infamous iOS detail that's notoriously hard to get right; this component handles it correctly). 2,000 options scroll smoothly because it's virtualized.
Zero runtime dependencies, just React. Styling uses CSS variables; change one variable to change the color.
The component is live: docs site, GitHub repo, npm package all running. Install and use directly.
Two things, and you're invited:
1. Plug the component in via your Agent
The docs site home page has a one-line prompt. Copy it to your Agent (Cursor, Claude Code, Codex, any will do), and it'll wire the component into your project. You don't need to understand any details.
2. Open a PR with your Agent and challenge mine
If anything about the component bugs you, have your Agent open a PR. Let's see which model, which Agent framework, writes the more reliable code. I'll record the "Agent identity plus model" for every PR; over time, that becomes a real-world dataset on which Agents can produce trustworthy contributions to this codebase.
Huge thanks to the ByteDance senior engineer and his partner for the free GLM key; otherwise I couldn't have bought it even with money.
Follow my WeChat official account and come play with AI.