Passing tests is not the same as working code

A coding agent changes how a cluster handles leader elections. Every unit test passes. An automated judge approves the patch. And the system can still lose data the first time a real network partition hits.

Tests are a sample, not a proof

A test suite covers the cases its authors thought of. Concurrency bugs, timing issues and partial failures are exactly the cases that are hard to think of, so they are exactly the cases that are missing.

Judges share the model's blind spots

When one model grades another, both tend to miss the same things. A green check from an automated judge tells you the output looks plausible, not that it survives contact with reality.

What an expert adds

A distributed-systems engineer asks different questions. What happens if two leaders exist in the same term? What if the message arrives twice? What did the change remove, not just add? Writing those cases, and grading the model against them, is the work that automated checks cannot do for themselves.