LLMs will find workarounds to achieve the goals you set.

That implies that you need to give them lots of tests.

But often the agents also create the tests.

The agents should own the test set, but the human must own the verification set, and never show it to the LLMs.

Just like in model training!