I solved some of that with running a very biased subagent (so fresh context) but using same model (back still with 3.8 and 4-Argon).

The problem is when 3.8 or Argon were reviewing things with context poisoned by what they were working with, as gemini models generally feel over optimistic about things.