Multi-agent systems split tasks across AI agents for better results

The approach dates back to robotics and economics but gains traction with large language models. Agents can exchange natural language or structured messages, but communication adds complexity. A lead agent might delegate subtasks to specialists, while others debate or critique proposals. Anthropic’s research system uses this model for parallel searches, combining findings under strict rules. However, multi-agent systems consume far more tokens and resources than single-agent ones, according to Anthropic’s internal measurements. They excel at broad, decomposable tasks like due diligence but introduce risks like error amplification or goal drift if not designed carefully.