The key is separating who writes the code from who defines the evidence—if one agent controls both implementation and acceptance tests, green checks can still certify the wrong thing.
how are you actually handling that separation in practice right now, different agent for the tests, or still mostly manual?
I am enforcing a separation by freezing the intent before the code is written and verifying the evidence afterward. This is important because the intents themselves can be modified by the agent during development, which can make everything appear successful at the end. I am developing Shipmoor, an agent verification loop to address this issue, and I would love to hear your opinion on it.