AI writes code fast — but who keeps it maintainable? An open-source bet on metadata

cpengc1984 — Fri, 05 Jun 2026 06:22:27 +0000

In 2026, AI writes code at incredible speed. Cursor and Claude Code agents read whole repos, the CLI became the new dev infrastructure, and "a team of agents working together" is the story of the year.

But there's an uncomfortable truth: AI writes fast, yet enterprise apps still fail to ship — maintainability gets worse, not better. Why?

Speed without "scale" (standards) is just faster tech debt

For a toy app, letting an AI agent one-shot the code is great. For an enterprise system, you need:

Maintainability — survive the 7th round of requirement changes
Evolvability — still extensible after 5 years
Deep integration + self-hosting + audit + fine-grained permissions

The problem with AI agents writing code: they emit a wall of one-shot code that immediately becomes a second source of truth alongside your project. The next agent run overwrites your hand edits; multiple agents each write their own thing. The faster it goes, the faster the debt piles up.

AI gave us speed. Nobody gave us scale — the standards, boundaries, and order. And enterprise apps die on exactly that.

The bet: make AI work on metadata, not on code

Oinone is an open-source, 100% metadata/model-driven low-code framework. The bet: data models, UI, permissions, workflows, AND the AI's output all live in one shared metadata model.

So an AI agent doesn't write throwaway code — it writes into the same metadata the framework and human developers already operate on:

What the AI changed is a structured metadata diff — reviewable, revertible, not 500 lines to read by hand
Multiple agents collaborate on one model, not divergent copies
Because metadata is compact, AI-coding token usage drops ~60% in our benchmarks

AI for speed, the framework for scale. That's the whole idea.

Try it (one command, ~5 min, no signup)

curl -L https://gitee.com/oinone/oinone-docker-shared/raw/master/oinone/docker-compose.yml -o docker-compose.yml
docker compose -p oinone up -d
# open http://127.0.0.1:88   user: admin   password: admin

Then have the AI generate an app from a sentence, and look at what it produces — a metadata diff, not a pile of code. That's the difference between "AI-native" and "a low-code tool with a chatbot bolted on."

Stack: Java backend + TypeScript frontend, AGPL-3.0 (genuinely open source, the framework you run is the framework that's public). Self-hostable; in production at large enterprises.

Honest about the rough edges

English docs are catching up; the polished 7.x live demo lands soon (the quickstart above is the real thing today). It's not for simple internal tools — use something lighter for those. It shines on complex, long-lived, self-hosted enterprise systems.

If the "metadata as the single source of truth for AI + humans" idea resonates, a ⭐ on GitHub (or Gitee) helps more developers find it. Happy to discuss the metadata model, the AGPL choice, and how it compares to Retool/Appsmith/Budibase in the comments.

DEV Community: cpengc1984

AI writes code fast — but who keeps it maintainable? An open-source bet on metadata

Speed without "scale" (standards) is just faster tech debt

The bet: make AI work on metadata, not on code

Try it (one command, ~5 min, no signup)

Honest about the rough edges