DeepSeek-R1 Grew Its Own Internal Debate Club — Nobody Asked It To

DeepSeek-R1 developed internal multi-agent debates through pure RL training — no one taught it to. Google researchers call this the 'Society of Thought.' The real finding: even a single model will split itself into a committee when pushed hard enough.

Kimi K2.5 Trains an Agent Commander with RL — SemiAnalysis Tests Show Claude Agent Teams Are Actually Slower and More Expensive

SemiAnalysis: Kimi K2.5's agent swarm uses an RL-trained 'orchestrator' (not prompt magic). Claude Agent Teams were slower, pricier, & scored lower. Multi-agent is shifting from 'prompt engineering' to 'distributed scheduling.'