An agent writes 10,000 lines before lunch. A human reviews a few hundred an hour, less if they’re being honest about it.
That gap widens with every model release. The obvious answer is tests.
Shift left just relocates the bottleneck
Shift left doesn’t close the gap, it just moves it. Now humans are hand-maintaining a test suite that’s growing as fast as the code.
Same arithmetic, different file.
Worse: when the same agent writes the code and the tests, the tests inherit its blind spots. You haven’t created an independent verification layer. You’ve asked the same system to grade its own homework.
The fix is shifting out
An autonomous QA layer that sits outside the thing it’s testing.
Independent context, derived from the spec, not the implementation. Adversarial by design. Its job is to break the code, not confirm it works.
Humans review the spec. Agents write the code. An independent QA layer verifies the behavior.
Nobody reads 10,000 lines.
