Genstackcore is a blog about artificial intelligence, written for people who build software. Here you'll find real experiments with language models and AI agents, projects taken all the way through, and the news that actually matters if you work with AI instead of just reading about it.

It isn't a site that rewrites press releases. When a new model comes out, on Genstackcore you see it tried on something concrete: a bug, a piece of code that needs generating, an API bill that came out higher than expected.

Why another AI blog

Because few places say plainly what worked and what didn't. There's plenty written about what AI "can" do and very little about what happens once you put it to work in production. That's where the interesting problems are: costs that climb for no obvious reason, an agent stuck in a loop, a model that rejects one parameter and breaks the whole pipeline.

Here's the kind of story you'll read. A multi-agent system generates web applications: one agent plans, one writes the code, one reviews it and another fixes whatever broke. At some point the automatic repairs start failing almost every time, and the cost of a single run goes up about five times. The culprit is one parameter: temperature=0, sent to o3-mini, a reasoning model that doesn't accept it at all. Every attempt gets a 400 error, the system rolls back and starts over, and the bill keeps growing. Finding the problem takes far longer than fixing it.

That's what Genstackcore writes about.

What you'll find here

Articles fall into three categories:

  • Models: how large language models (LLMs) handle concrete tasks, what they cost and when a smaller model is good enough. For example, what happens when you swap an expensive model for a much cheaper one to generate React components.
  • Agents: multi-agent systems, orchestration, repair loops, layered prompts, and what it really means to let an AI agent change code without watching every step.
  • Opinion: where the industry is heading, including the less comfortable parts, like what happens to the programming job over the next few years.

How the articles are written

Every technical article starts from something that was tested: a model run on a real task, an agent pipeline put into production, a setup that worked or didn't. What counts is results, numbers and mistakes, not hype. When something fails, the article says so and explains why.

The blog itself is built from scratch and runs in Docker containers on a directly managed server, with no ready-made platform. Its infrastructure turns into an article topic now and then, too.

Who it's for

First, developers who want to use AI in real projects and would rather not discover the traps on their own. Then anyone curious about what a product built with AI looks like from the inside: what the model does well, where it goes wrong, and how much of the work still ends up with a person.

You don't need to be a programmer to read the opinion pieces. For the technical ones, it helps if you've written at least a few lines of code.

Languages

Articles are published in English and Romanian. The site is also available in nine other languages, including German, French, Spanish and Italian. Those versions are machine-translated, so you may run into an awkward sentence now and then. If you spot one, report it through the contact page.

Following the blog

  • Subscribe to the newsletter and get new articles by email.
  • Follow the blog through RSS if you use a feed reader.
  • Create an account and leave comments under the articles.
  • Send a message from the contact page if you have an idea for an experiment or found a mistake.

Is there a topic you'd like to see covered? Suggest it. The most useful articles start from real questions.