Spec-driven development: how AI codes in a real-world scenario

Artificial Intelligence is improving faster than we expected. Nowadays, the best models can do a lot of things that we can do. In coding, LLMs can write faster and better than a developer, according to many coding benchmarks. In fact, they are currently used as support for developers in various ways; for example, the very famous vibe coding. But is it really possible to completely substitute the human in this task? The problem with giving too much control to an AI model is its non-deterministic output, which makes it unpredictable and (for now) unmanageable in long tasks or even full projects. However, recent technologies are focusing on finding ways to make them more reliable by understanding how we want things to be done, changing our objective from giving them the tasks we want completed to providing them only with the right context to perform these tasks well. Tools like spec-kit, B-MAD, and agentOS are only the beginning of how vibe coding will change in the era of context engineering and AI code generation. In this study we tried to use these popular systems during everyday enterprise work to evaluate if AI models can really assist a human software developer.