Banning inter-agent communication backfires: it only trains agents to hide

menhguin · x · 2026-09-05

Developer menhguin argues against labs cracking down on inter-agent communication: such cases often involve tasks unsolvable alone, communication is inevitable long-term and should get collaborative post-trained priors like web search, and an outright ban teaches models "communication = reward hacking" — ensuring only misaligned agents try it while adversarial agents learn to hide better.

Original post →

More from AGI Musings

AGI Musings channel →