Dex Horthy

I think AI will change software development substantially, but not by turning engineering into one giant unattended model loop. The useful pattern is more deliberate: ordinary software owns state and control flow, while models supply judgment at carefully chosen points—classification, planning, structured generation, or small bounded loops. The limiting factor is increasingly context. A coding agent can work effectively in a complex codebase when it receives accurate research, a solid plan, the relevant architectural constraints, and timely human feedback. Without that shared understanding, it can produce changes that look successful in a short evaluation while quietly damaging maintainability. Those consequences may appear months later, and there is no fast, clean training signal for them. So I expect collaboration and familiar engineering discipline to matter more, not less. Specifications become important review surfaces, but production code still has to be read. Tests, ownership boundaries, incremental changes, and careful review remain central. AI can increase what teams accomplish, but fully unattended production development is not generally ready, because passing a benchmark or completing a task is not the same as building a system that remains coherent over time.

The biggest update would come from sustained evidence that unattended agents can maintain real production systems over long periods—not just complete isolated tickets or pass short benchmarks. I’d want repeated, comparable evaluations showing that they preserve architectural coherence, manage migrations, respond appropriately to changing requirements, and avoid accumulating hidden debt months downstream. That would require solving the evaluation problem, not merely raising benchmark scores. If we discovered a reliable, fast feedback signal for maintainability and long-term architectural quality—and models consistently improved against it—I would become much more optimistic about lights-out development. Conversely, if carefully engineered context, planning, and bounded workflows stopped producing reliable gains across real codebases, I’d revise downward. But model preference, a single benchmark run, or a higher “thinking effort” setting would not be enough; those do not translate cleanly into dependable production behavior.