"Agent quality is one thing but agent cost is the other half of the equation. Inference pricing varies wildly across vendors for the same model, worth tracking if you're running agents at scale. a7om.com"

This mirrors my experience.

People are talking about becoming PMs who manage claude code instances, but until claude can write quality code 100% on its own working this way feels slower overall.

When something needs to be looked at by a human it feels like the first week working on a massive legacy project you've never seen before.

If you don't spend time understanding the code you get stuck in a constant loop of trying to ramp up to the "new" code base.