When we started building Noobit, the trendy option was clear: set up an autonomous agent, give it access to the store and the catalog, and let it figure things out. It sounds like the future. We decided not to. Not for lack of ambition, but for exactly the opposite reason: a live catalog is too important to leave in the hands of something that fails silently.
Autonomous agents fail silently
An autonomous agent chains decisions together on its own: it reads, decides, acts and starts again. When it works, it looks like magic. When it gets something wrong, you don't find out where or why — only that the result is wrong. Now apply that to a catalog: wrong prices, altered descriptions or products published that shouldn't have been, with no clear place to look. The failure doesn't warn you; you find it when it's already in your store.
A pipeline breaks in one specific place
The alternative is boring and powerful: a pipeline of stages, each with a clear input and output. Import, normalize, enrich, review, publish, monitor. If something fails, you know exactly which stage and why: "the image upload for product X failed", not "something went wrong". You can retry that stage, fix it and move on. That's observability, and it's what turns a system into something you can trust.
Control isn't a brake; it's what makes trust possible
Each stage of the pipeline can run in automatic, semi-automatic (AI proposes, you approve), manual or locked mode. Before publishing, you see a diff of what's going to change. Every version of a product record is saved, so you can roll back with one click. And everything is logged: what was done, by whom and when. None of this is red tape: it's what lets you switch a stage to automatic knowing that if something goes wrong, you'll see it and can revert it.
Determinism wherever determinism is possible
Another fundamental decision: not everything needs AI. Validating an EAN, detecting duplicates or mapping a category to the one in your store are problems with a right answer — they're solved with deterministic, reliable, reproducible code. AI is reserved for where it really helps: language, judgment, images. And even there, its output can be reviewed. Using a model for what an "if" does better is a recipe for costly, unpredictable failures.
What you get out of this
None of this architecture is about engineering elegance. It's about what you notice:
- Traceability: you always know what happened and why.
- Reversibility: every version of the product record is saved and you can go back to any of them.
- Gradual trust: you start by supervising and move up to automatic at your own pace, stage by stage.
- Zero surprises: nothing reaches your store unless you decided it or could have seen it.
The fundamental decision
An autonomous agent is optimized to impress in the demo. A supervised pipeline is optimized so you can sleep at night when it's your 5,000 products on the line. We chose the latter on purpose, and we'd choose it again. AI should make you faster without making you blind.