Why a factory.
Why now.
The market thesis behind OOretz Factory — four trends that make 2026 the right year, and why a web-only ceiling exists for AI app builders. Today we ship web too; the bet is to be first across the ceiling.
Different from the manifesto (our beliefs). This is the market case (the timing).
Frontier models cleared the "build a real app" bar
Three years ago the autonomous-build category produced toy demos. In 2026 it produces real apps.
Claude Sonnet 4.6, GPT-5, Gemini 2.5, and DeepSeek V3 routinely write thousands of lines of compilable C# / TypeScript / Python / XAML in a single conversation. They survive multi-step reasoning. They follow architectural constraints handed to them as system prompts. They self-correct when shown a compile error.
The frontier is no longer "can a model write a working function?" The frontier is "can a model coordinate a build pipeline that ships an installer?" That coordination layer is what the factory adds on top of the models. The models alone are now strong enough that the layer is worth building.
Benchmark progress on SWE-bench Verified: 6% in 2023 → 71% in late 2025 (Anthropic public eval).
The web-only ceiling stops being acceptable
AI app builders that ship only deployable web apps are leaving half the economy on the table.
Lovable, v0, Bolt, Replit Agent — every one of them is excellent at what they do. They produce beautiful web demos and deployable Next.js apps. The ceiling is "deployable web app."
But the software running the actual economy isn't web. It's the desktop app at your bookkeeper's office that prints MICR checks. It's the WPF tool the warehouse uses to scan barcodes. It's the Win32 program your lawyer pastes deal terms into. SaaS hasn't reached these workflows in 15 years and isn't about to.
Buyers in those segments don't care about deployable Next.js. They care about an installer. The first autonomous factory that produces an installer wins those segments.
The labor math finally favors automation
Custom in-house software used to require a developer. In 2026 the math shifts.
A US software engineer costs roughly $200,000 fully loaded per year. A custom internal tool used to require 2-3 engineer-months to build. Call it $50,000-$100,000 of labor.
Generating that same tool through an autonomous factory costs $50-500 in AI tokens, takes 30 minutes wall-clock, and produces an artifact with hash-verified proof. That's a 100-1000× labor compression.
Compression of that magnitude doesn't just speed up existing buyers — it creates a new tier of buyer. Small businesses that couldn't justify a custom app at $50,000 can justify one at $50. That tier is much, much larger than the developer market.
Compliance + audit makes "AI wrote it" credible
Two years ago, "AI wrote this" was a liability. In 2026 it can be a credential — if you do the proof right.
Auditors used to balk at AI-generated code: "but how do you know it does what it says?" The honest answer used to be: you don't.
In 2026, the components a regulated buyer needs are starting to land. Hash-chained audit logs with actor_is_agent columns. Structured proof packs per build (compile attempts, repair cycles, the actual artifact). Provider-side BAAs and DPAs from the major LLM vendors. SOC 2 with controls scoped to AI inference paths. Some of this we ship; some is on our roadmap. None of it is fully solved industry-wide yet — but the trajectory is clear.
The factory is built around this. Every passing artifact has a SHA-256 hash and a downloadable URL. Every job has a proof pack. /factory/proof shows the live list — passes AND failures — so an auditor can audit the audit. SOC 2 Type II in progress. Per-lifecycle audit-event emit is not yet wired into the orchestrator (in progress). "AI wrote it" becomes credible when the proof trail is real, not when the marketing copy says it is.
The bet, stated bluntly.
- →Frontier models are good enough to do the coding. The bottleneck is the build pipeline around them, not the models themselves.
- →Web-only builders take the easy half of the market. The hard half — installable native software — is wide open and worth more. Today we ship web (the same surface they do, with a smoke pass bar they don’t); the bet is to get verified desktop and mobile shipping before they do.
- →Labor compression from $50K to $50 unlocks a buyer tier that didn’t exist before. That tier is the actual prize.
- →Auditable AI is now possible. The first factory whose proof trail is more rigorous than a human team’s wins the regulated buyers.
If you think any of this is wrong, write to us at [email protected]. We’ll argue, kindly.