I think "how much coding?" is the wrong unit.

The better question is: how much independent judgment do you need before delegating implementation?

You should learn enough to:

- turn vague requirements into interfaces, invariants, and tests - read a diff and explain what it actually does - debug without asking the same system that wrote the bug - recognize when the architecture is wrong rather than continuing to patch symptoms - reason about state, concurrency, failure, security, and cost

That probably means learning one language deeply, plus basic data structures, databases, networking, operating systems, version control, and debugging. Not because you will hand-type every line forever, but because those are the mental models required to supervise an agent.

The curriculum should change, not shrink. Less time memorizing syntax; more time decomposing problems, reading unfamiliar code, testing behavior, instrumenting systems, and understanding failures.

I would use AI early, but periodically remove it. If you cannot build and debug a modest program without the agent, you do not yet know whether it is accelerating you or substituting for understanding.

A practical rule: on’t merge code you can’t explain, and don’t deploy a system whose important failure modes you can’t enumerate.