The first time I watched a coding agent finish a task in under a minute and then spent the next hour figuring out whether the change actually worked, I understood what the productivity numbers were hiding.
Generating the code was fast. Validating that it worked was not. It was slow because the change touched a service that talked to four other services, and there was no way for the agent to exercise that path without me standing up an environment, redeploying, and running the request myself. By the time I verified the change worked, I had spent more time validating than the agent had spent writing.