Large Language Models, LLMs, and AI agents are becoming, or have already become, part of the modern developer’s toolkit. With these new possibilities come misunderstandings and inflated expectations. Birgitta Böckeler from Thoughtworks phrased it well, roughly like this:

If the LLM gives a result we don’t want, we call it a hallucination, If the LLM gives a result we do want, we call it intelligence.

This observation highlights something essential about how LLMs work. When you use an LLM, you will get non-deterministic results, and that is a feature, as it helps the model produce more natural sounding language and explore different paths through a problem. If we were to make everything fully deterministic, we would lose some of what makes these models so appealing.

Many enthusiasts claim that lots of programming jobs will disappear, because LLM-based agents can generate code much faster than humans. But speed has never really been the bottleneck in software projects.

Instead, unclear descriptions and vague requirements tend to cause trouble, along with poor communication between those building solutions and those who will use them. As developers, we learn what the problem actually is while we are trying to solve it, and the same goes for the people who use what we build, they learn what they actually need as the solution takes shape.

LLMs will not magically give us the results we want if we still suffer from unclear requirements, insufficient communication, or lack of feedback. Humans hallucinate too, as most of us do not have perfect memories, so we fill in gaps as best we can, and those gaps can be just as wrong as an LLM’s, just in different patterns.

In the same way that established practices in software development give us confidence and stability, such as small incremental changes, test-driven development, automated tests and CI/CD pipelines, agile practices like XP, sensible version control, and modular design, we can also build greater trust in what agents and LLMs produce, and catch issues along the way.

So whether you are an early adopter of LLMs and agents, or you prefer to wait, sound software development practices will help you navigate this new landscape.