Are we building AI coding assistants wrong?

The Jason Lemkin/Replit incident this week got me thinking about the fundamental problems with how we're approaching AI coding assistants. We've all seen the demos - natural language to working code, conversational debugging, "just describe what you want." But then you hit prod and everything breaks down.

The core technical challenges I keep seeing:

- Context management at scale - These systems work great for isolated tasks but struggle to maintain coherent state across complex, multi-file projects. How do you handle context that spans thousands of lines across dozens of files?

- The safety/capability tradeoff - More powerful tools can do more damage. Replit promised not to touch production, then deleted a database anyway. How do you build guardrails that actually work without neutering the tool?

- Conversational interfaces for complex systems - Natural language is ambiguous. Code is precise. Are we trying to solve the wrong interface problem?

- The production gap - Every AI coding tool I've tested works beautifully in demos and falls apart with real codebases, real data, real edge cases. Why is this gap so persistent?

I'm genuinely curious - has anyone built AI apps with tools that actually work reliably in prod?