When a prompt that worked perfectly stops working, engineers assume they broke something. But LLMs are non-deterministic: the same input produces different outputs. Continuous evaluation, not perfect prompts, is how you build reliable AI products.
What makes AI product development so damn hard? It's because you're building with non deterministic pipelines. Here's the difference. Most of your stack APIs, functions, databases. It's deterministic. You're used to getting a product requirements document. You get some input, you get an output. There's ground truth. You could test it. LLMs, not so much. Two identical prompts can give you totally different answers because the model's path isn't fixed, it's probabilistic. That's non determinism and it means it worked once. Doesn't mean it just works every time. One client I saw built a great agent demo but it actually failed the same task in prod in a different environment. Why is that? It's because of non determinism and that's why evals matter. We're not just shaping the code. You need to get to a continuous improvement cycle of scientifically and systematically improving your eval process so that the subjectivity is also improved in the user experience with your AI or LLM applications.
