agent-swarms
1 articles
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.'