This is a good reminder that tooling choices compound over time. The short-term speedup matters less than whether the next maintainer can still reason about the system.
The distinction between 'vibe coding' and 'agentic engineering' is important. In my experience, the key difference is whether you're reviewing and understanding the code the agent produces. When I use coding agents for non-trivial tasks, I always review the diff before committing — that's the engineering part. The danger is when people skip that step and just trust the output.
I like the small-surface-area approach. The question I’d use to evaluate this is how well the harness records/replays tool calls and failure modes, since that is where debugging agent behavior usually gets messy.