This is a talk given at the Wallenberg Advanced Scientific Forum on the theme Foundations of NeuroSymbolic Artificial Intelligence: https://wasp-sweden.org/wallenberg-advanced-scientific-forum-2026/ Large language models are empirically strong (if flawed) at conventional language processing. They summarise, translate, generate, and converse with accuracy rates that would was hard to engineer a few years ago. But there is something they are not good at: reasoning (and planning), at least in a reliable and robust manner. This is the motivation for "neurosymbolic executors". The pitch is simple. Let the LLM do what it does well — understand natural language, parse problems, generate structured output. Then hand the actual reasoning off to a symbolic system that can do it correctly, transparently, and verifiably.