This is a very relevant topic. One thing I’m curious about is how you handle non-determinism in agent evals. Even with the same input and prompt, different models or model versions can give different results. Do you compare only the final answer, or also the steps the agent took to get there?

I like this approach. AI agent output can look good in one example, but fail in the full workflow. Treating prompt and agent changes like code changes makes sense to me. You need baseline tests, edge cases, and regression checks before trusting it in production.

Data cleaning is not the most exciting part ) I would be interested to see this on a concrete case. Have you had a real example where bad or messy data quietly broke the AI output, even though the model itself looked fine?

This is very close to how I see the change too. AI can already write a lot of code, but that does not make it a product engineer. The important part is understanding the problem, the constraints, the risks, and knowing when the result is good enough. In my work, the human role is moving more toward owning the context and making the final decisions.

Great question. I think logs for AI agents need to be much more structured than normal logs. Trace IDs, clear error context, input values, expected result, actual result, and safe steps to reproduce would help a lot. Without that, the agent may just make guesses from incomplete information.

I agree that strict typing is becoming more important, especially when AI tools work with the codebase. Good types make the code easier to understand and safer to change. But I still think some documentation is needed to explain why things are built that way, not only what is allowed by the types.

Nice idea. Managing many platforms can easily become a separate job. For me the most useful part would be not just scheduling posts, but helping me understand what actually needs attention today and what can wait.

Nice honest story. Five years is a long time, but I think BA skills often pay off slowly. They help you ask better questions and understand what the business really needs. This feels even more important now, when AI can write code, but someone still has to decide what should be built.

Utility props look useful because they make styling simpler when you build for both mobile and web. My main question is migration: how hard is it to move a real app with many custom NativeBase components to gluestack without rewriting too much?

Stack Overflow helped developers a lot for years, but with AI tools I honestly don’t remember the last time I opened it directly 😊 The open question is where expert Q&A fits now. AI can work with context, while human experts bring real-world experience and judgment, but that experience also needs to be relevant to the specific problem. How do you think expert Q&A platforms should handle that matching and validation problem?