Engineering after code generation

Generating more code is an intermediate result. Engineering leaders still need to know whether the software preserves the right behavior, whether their teams can diagnose it, and whether newer engineers are learning enough to take responsibility for it. Read these pieces together as an argument about the work around implementation.

Preserve an independent test of intent

If one agent writes a requirement, implements it, and generates a test that repeats its own interpretation, agreement among those artifacts can be misleading. Use evidence that exists independently of the implementation: domain examples, accepted transactions, established interfaces, and judgments from people who operate the service. Keep expected behavior open to challenge.

Migrate behavior, then retire the dependency

Legacy replacement becomes difficult at the connections: reports, batch windows, shared data, and downstream assumptions. A working new interface is only one part of a migration. Identify the behavior that must survive and the behavior worth changing; those are separate decisions. Plan how the organization will compare results and recover before it stops relying on the old system.

Make judgment teachable

An apprenticeship should expose new engineers to decisions as well as tools. Have them explain a failed test, trace a request across a boundary, or defend a rollback plan with an experienced colleague. Review can become a place where systems knowledge is transferred, provided the team has time to teach and a clear owner for the resulting work.

Questions for discussion

Who supplied the expected result, independently of the generated implementation?
Which hidden dependency could invalidate a successful demo?
What can a new engineer explain and operate without the agent?

Reading list

Who verifies the verifier?
Examine Karpathy’s argument about verifiability and its limits in delivery.
Before the rewrite, prove the behavior
Follow a replacement through behavioral contracts, reconciliation, and cutover.
The apprenticeship AI does not replace
Consider an engineering team designed to develop judgment alongside AI fluency.