The SonarQube anecdote shows well that the useful skill is detecting that something is wrong, not writing the fix. One thing I'd add to the "just-in-time depth" idea is making those detectors explicit and independent of the agent: complexity budgets, dependency diff alerts, mutation testing on critical modules, and a short architecture-decision file the agent must update when it changes a boundary. They act as the alarms that tell you when to dive. The training question gets harder, though. Do you think juniors should deliberately debug agent-written code in a sandbox with seeded faults, so they build the pattern recognition that used to come from writing it by hand?