Imagine a population of machine agents. Each might be strong on certain tasks but fundamentally limited: partial tools, partial observations, finite context, bounded compute. How can these agents self-orchestrate and self-evolve into stronger collective intelligence to solve tasks beyond any single agent's capability? Instead of designing the multi-agent system... itself, we propose designing the incentives that govern it. We put agents in an economy. They compete, trade, get wealthy, go bankrupt, and mutate, forming an alive society where coordination and adaptation automatically emerge in a decentralized manner.show more

1/n We introduce Economy of Minds (EoM), which has two coupled processes: Planning: how agents coordinate within a task. Adaptation: how the society evolves across tasks.

2/n For planning, each agent just has two local components: 1. a wake-up condition: when should I act? 2. an action policy: what should I do? At each step, every agent independently decides whether to wake up in the current state.

3/n The awakened agents then participate in an auction. Each eligible agent submits a bid. The highest bidder wins the right to act, executes its action, and advances the environment to the next state. No central controller decides who should act next. The market allocates control.

4/n After acting, the winner pays its bid to the previous winning agent. These transactions create a decentralized credit assignment mechanism. Agents are rewarded not only for producing final solutions, but also for creating intermediate states that enable future progress. When downstream agents are willing to pay to continue from a state, value flows back to the agents that helped create it. As a result, useful reasoning steps, tool calls, design edits, and code modifications can all be rewarded through the market.

5/n For adaptation, the society evolves through wealth. Agents that repeatedly contribute to successful trajectories naturally accumulate wealth. These wealthy agents are selected for exploitation: they produce mutated descendants that inherit and refine useful behaviors. Successful strategies are preserved and diversified.

6/n Agents that fail lose wealth. They may act in unproductive states, make poor decisions, or simply fail to attract downstream continuation. When their wealth becomes negative, they go bankrupt and are removed. New agents are injected through exploration: failed agents amending themselves by exploring new strategies.

7/n So EoM does not train a central planner. It does not prescribe a communication graph. It does not require global awareness. Each agent only knows when to wake up and how to act. The society-level intelligence emerges from local auctions, payments, wealth accumulation, bankruptcy, and mutation.

8/n We instantiate EoM with language agents on five domains: 📐 mathematical reasoning 💰 financial research 🔬 scientific research ⚙️ accelerator design ☁️ distributed-system optimization Across all domains, we start from weak or partial agents and ask whether the economy can turn them into a stronger system.

9/n On MATH, partial agents are initialized as planner, executor, and verifier agents with limited output budgets (128 output tokens). Individually, they are weak. But after economic evolution, Llama-3.1-8B agents could improve from 15.9% → 57.0%, even outperforming the corresponding complete-agent baseline (i.e. w/o constraints).

10/n On Finance-Agent-Bench, each partial agent has access to only one tool. No individual agent can solve the full research task alone. Through economic coordination, EoM improves from 45.0% → 60.0%, outperforming complete-agent baselines.

11/n On scientific research (Frontier-Science-Research), EoM evolves reusable scientific reasoning routines. Agents learn to decompose problems, identify governing principles, check constraints, verify equations, and transfer these patterns across domains.

12/n On accelerator design, EoM searches for better hardware mappings. The economy discovers useful design strategies and improves average EDP to 39.3, outperforming both a same-backbone complete agent and a strong domain-specific baseline.

13/n On distributed-system optimization task, EoM evolves coding agents that improve a Cloudcast program. It reaches a best cost of 657, compared with 930 for OpenEvolve. The society learns when to read, edit, build, evaluate, and finalize through market-selected workflows.

14/n We also observe emergent specialization. Even when a complete generalist with access to all tools is added, it does not monopolize the economy. Specialists survive because they become locally precise and useful. The market rewards contextual value, not broad capability alone.

15/n We hope EoM could suggest a different way to build multi-agent systems. Instead of manually designing every role, workflow, and communication protocol, we can design incentives, from which coordination, specialization, and adaptation automatically emerge.

16/n Huge thanks to amazing collaborators @Huangyu58589918, @ao_qu18465, Chenyu Wang, Yu Yao, Han Zheng, Kushal Chattopadhyay, @Kevin_GuoweiXu, Zihan Wang, Weirui Ye and advisors Vijay Janapa Reddi, Ju Li, @pliang279 @hima_lakkaraju @ShamKakade6 @du_yilun! 📄 Paper: 🌐 Webpage: 💻 Code:

Interesting, may I ask how does the agent know how much to bet? Are there optimal bets for each agent?

hi! really enjoyed your thoughts here… very inspiring. we’re building a multi agent product and i’d love to get your feedback if you’re interested. we’re still in beta, but we’re looking for thoughtful early users. team is from google and stanford :) open to giving it a try?

My contribution toward this architecture, coordinated memory: