Decoupling makes the isolated change easier while coordinating the change that affects multiple layers harder.
People used coupled frameworks to make the layer spanning changes easier. AI is not great at such changes either but it can put in the work to properly split the change it is trying to make into separate per-layer changes ahead of time. So I think in general you are right that for AI it's better to have things very decoupled.
It's still worth having things that ensure consistency on layer boundaries, like for example using backend types (dtos) to generate fronend types so incompatibility doesn't go undetected.
People used coupled frameworks to make the layer spanning changes easier. AI is not great at such changes either but it can put in the work to properly split the change it is trying to make into separate per-layer changes ahead of time. So I think in general you are right that for AI it's better to have things very decoupled.
It's still worth having things that ensure consistency on layer boundaries, like for example using backend types (dtos) to generate fronend types so incompatibility doesn't go undetected.