24 GPT-6 Astra agents attack each other's answers until one survives
The author had 24 agents produce 41 competing decisions and then attack each other's results, narrowing down to one answer, concluding the system got stronger by eliminating bad answers rather than piling up ideas.
i gave GPT-6 Astra 24 agents one problem they produced 41 competing decisions before a single answer was allowed to survive then i made every agent attack the result at first it looked completely unnecessary some agents reached the same conclusion through different evidence others disagreed on tiny assumptions that changed everything downstream a few found answers that looked stronger than the one i would have picked myself so i stopped treating disagreement like noise each decision had to survive another pass one agent challenged the source another challenged the logic another looked for the hidden assumption another tried to break the result with a completely different path weak answers started disappearing fast then the interesting part happened the system stopped improving by adding more ideas it improved by killing bad ones earlier 41 decisions became 17 17 became 6 6 became 2 and those final 2 kept attacking each other until only one was still standing the final answer wasn’t the smartest sounding one it was the one that had survived the most pressure that changed how i think about GPT-6 Astra i used to optimize for better first answers now i’m much more interested in building systems where bad answers simply don’t make it to the end
This post quotes another post: x.com/lummox_eth/status/2100288151897981255