Economy of Minds: Emerging Multi-Agent Intelligence with Economic Interactions

Economy of Minds, developed by researchers affiliated with Harvard, MIT, 2077AI, and the Kempner Institute, proposes a decentralized multi-agent framework in which LLM-based agents compete through auctions, exchange payments via a bucket-brigade rule, and evolve through wealth-based selection rather than centralized orchestration. Each agent is only a local wake-up predicate and policy, with bids frozen at introduction and wealth updated by downstream rewards, so credit assignment emerges from market dynamics instead of explicit global supervision. Across five agentic tasks, the system turns weak partial specialists into stronger collective systems: on MATH, Llama-3.1-8B improves from 15.9% to 57.0% and Gemma-2-9B from 4.2% to 45.1%; on Finance-Agent-Bench it rises from 45.0% to 60.0% after 30 training tasks; on FrontierScience-Research it reaches 20.0% best-run accuracy; on GEMMINI accelerator design it cuts average EDP to 39.3 versus 80.2 for DOSA; and on Cloudcast it achieves a best cost of 657 versus 930 for OpenEvolve. Ablations show the gains depend on the full economic loop: removing exploration drops Finance-Agent-Bench mean accuracy to 26.0%, while removing auctions reduces it to 48.0%. The theoretical analysis links bids to expected payoff, shows outcome-only reward can suffice once the market selects near-optimal specialists, and characterizes bucket-brigade payments as Bellman-like and Shapley-like credit signals.