Research map / Multi-agent systems and games / LLM agents
Economy of Minds: Emerging Multi-Agent Intelligence with Economic Interactions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Economy of Minds (EoM), where LLM agent populations compete via bid-based auctions for control, pay each other bucket-brigade style, and evolve through wealth-driven exploitation and exploration, yielding emergent multi-agent workflows across five agentic tasks.
Problem
How a population of individually weak, partial LLM agents can self-orchestrate into stronger collective intelligence without a central orchestrator or explicit communication protocols.
Approach
Replaces central orchestration with a Hayek-machine-style economy: agents bid in auctions for control, bid payments propagate credit backward along the action chain, rent and bankruptcy remove unproductive agents, and wealthy agents are mutated via their prompt generators while bankrupt ones are amended. Only prompts evolve; the backbone stays frozen.
Key results
- On MATH, EoM raises Llama-3.1-8B partial agents from 15.9% to 57.0%, exceeding the complete-agent baseline of 51.9%.
- On accelerator design, EoM reaches average EDP 39.3 versus 43.1 for a complete ReAct agent and 80.2 for DOSA.
- On Finance-Agent-Bench, EoM rises from 45.0% to 60.0%, above Multi-Agent Debate (50.0%), ReAct (45.0%), and GEA (50.0%).
Surprising findings
- A strong generalist agent with access to all tools did not monopolize the economy; specialists continued to grow and the generalist contracted back to a single agent.
- Removing exploitation on Finance-Agent-Bench lowered mean accuracy to 33.5 while best-run accuracy stayed at 60.0, and the easy-to-hard curriculum was not strictly required for improvement since reversed order still improved to about 47%.
Limitations
- Adaptation occurs only in prompt space with a frozen backbone, which may limit capability growth on tasks needing new skills.
- Theoretical results rely on strong assumptions (recurrent contexts, calibrated bids, eligible near-optimal specialists).
Experimental setup
Datasets: MATH Finance-Agent-Bench FrontierScience-Research Gemmini benchmark (ResNet-50 kernels) Cloudcast (ADRS)
Benchmarks: MATH Finance-Agent-Bench FrontierScience-Research Gemmini ResNet-50 kernels Cloudcast
Metrics: accuracy mean and best-run accuracy energy-delay product total cost
Baselines: ReAct GEA OpenEvolve Multi-Agent Debate DOSA
Builds on: Hayek machine bucket brigade OpenEvolve ReAct Multi-Agent Debate
Closest papers by meaning
- When AI Agents Compete for Jobs: Strategic Capabilities and Economic Dynamics of AI Labour Markets (ICML 2026 · similarity 0.87)
- Scaling Small Agents Through Strategy Auctions (ICML 2026 · similarity 0.86)
- HiveMind: Contribution-Guided Online Prompt Optimization of LLM Multi-Agent Systems (AAAI 2026 · similarity 0.86)
- EconGym: A Scalable AI Testbed with Diverse Economic Tasks (NeurIPS 2025 · similarity 0.85)
- Shapley-Coop: Credit Assignment for Emergent Cooperation in Self-Interested LLM Agents (NeurIPS 2025 · similarity 0.85)
- AdaSociety: An Adaptive Environment with Social Structures for Multi-Agent Decision-Making (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.