One thing I became increasingly uncomfortable with was letting the AI architect decide how its own proposal should be challenged.
It was not because I thought the Architect was weak. Most of the time, it was capable, careful, and grounded in the source. But once it had formed the first coherent interpretation of a task, later reviews could still remain inside that interpretation.
Several agents could participate and the process could look independent, while the original framing continued to shape what everyone considered relevant. I eventually realised that choosing the critics also shaped where the reasoning was allowed to go.
Where the instinct came from
I first connected this to checks and balances. The point was not political, and it was not that an AI system might be morally corrupted by authority. What interested me was the structure: one part of a system should not define the rules, apply them, and judge its own application without resistance. Competence does not remove blind spots.
