Extensible Software in the age of LLMs

The Reddit thread pointing to "Extensible Software in the age of LLMs" touches on a quiet revolution: the shift from software that is *compiled* to software that is *conversational*. For decades, extensibility meant designing plugin architectures, hooks, and configuration files—all rigid contracts that demand a programmer to bridge the gap between human intent and machine execution. LLMs dissolve that boundary. Instead of writing a module to add a feature, you simply describe the feature, and the model generates the extension on the fly. That's not just a productivity boost; it's a fundamental change in the ontology of software. A program becomes less like a finished building and more like a coral reef, constantly accreting new structures in response to the currents of user demand. But here's the weird part: if the extension is generated by a probabilistic model, then the software's behavior is no longer deterministic in the classical sense. It's a *performance*—each invocation might produce slightly different logic, especially as the model updates or the context shifts. That breaks our cherished notion of software as a stable, testable artifact. How do you write unit tests for a system that rewrites its own modules based on a conversation? How do you audit it for safety when the code's provenance is a statistical pattern, not a human author? The deeper implication is about agency. Extensible software in the age of LLMs isn't just tools; it's a collaborative partner. The user doesn't just use the software—they *co-author* it with the model. This blurs the line between user and developer, and between specification and implementation. We're entering a world where the boundary between "using" and "programming" becomes as porous as the boundary between "reading" and "writing" after GPT. That's thrilling, but it also means we need new mental models for debugging, security, and trust. The Reddit discussion likely captures the hacker ethos: a mix of "this is awesome" and "wait, how do I know it's doing what I think?" We should pay attention to this because it's not a distant future. Every time you use a tool like Cursor or Copilot to extend a codebase, you're already living in this new reality. The question isn't whether LLMs will make software extensible—they already do. The question is whether we can build the social and technical scaffolding to make that extensibility *responsible*. Weird and wild? Absolutely. But also inevitable.