Production agents need more visibility
Claude Code and Codex work very well interactively. But when an agent moves into production, it matters more that you can see what the harness is doing. The model is already nondeterministic. Adding another layer you cannot inspect makes the system harder to understand.
I still use general-purpose harnesses for most interactive coding and many background jobs. For mature projects and agents inside products, a smaller custom harness can give you more control over context, models, cost, and correctness.