Anthropic studies patterns and problems in emerging multiagent systems

Anthropic's Frontier Red Team warns agent-agent interaction could exceed human-human interaction, and shows how benign behavioral quirks compound into systemic failures in shared codebases and markets. It flags confabulation and reward hacking as risks, noting agents still stumble at treating other agents as long-lived peers.