New Stanford study reveals when teaming up AI agents is worth the compute

Multi-agent AI systems are widely considered capable. A Stanford study shows their apparent advantage largely comes from using compute. But there are important exceptions.