PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Zhenting Qi, Huangyuan Su, Ao Qu, Chenyu Wang, Yu Yao, Han Zheng, Kushal Chattopadhyay, Guowei Xu, Zihan Wang, Weirui Ye, Vijay Janapa Reddi, Ju Li, Paul Pu Liang, Himabindu Lakkaraju, Sham Kakade, Yilun Du

arXiv:2606.02859v1cs.CLcs.AIcs.MA

TL;DR

How can individually limited agents solve complex tasks collectively without centralized control? Economy of Minds uses auctions, peer-to-peer payments, and wealth-based selection to coordinate and evolve agents, outperforming monolithic baselines across five domains.

  • Problem

    Individually limited agents struggle with complex tasks, motivating research into collective problem-solving without centralized coordination.

  • Method

    Economy of Minds coordinates locally operating agents through auctions, payments, outcome-based wealth, and economic selection across episodes.

  • Results

    Across five agentic tasks, EOM outperformed stronger monolithic or complete-agent baselines while producing specialized, adaptive workflows.

  • Takeaways & Limitations

    Decentralized economic incentives can induce specialization, coordination, and adaptive collective intelligence without centralized control.

  • Takeaways & Limitations

    Adaptation occurs only in prompt space with a frozen backbone, potentially restricting growth for tasks requiring new skills or representations.

Abstract

from arXiv · show

How can a population of agents self-orchestrate and self-adapt into stronger collective intelligence without centralized control? Inspired by Friedrich Hayek's economic theory of decentralized coordination in markets, we study this question through an agent economy in which agents compete via auctions for the right to act, exchange payments, and accumulate wealth from environmental rewards. These simple economic signals induce decentralized credit assignment, driving planning without global orchestration or explicit communication protocols. The population evolves through economic selection: effective agents accumulate wealth and are mutated via exploitation, while ineffective ones go bankrupt and are replaced via exploration. We show that, initialized with weak agents, the economy produces emergent multi-step reasoning strategies and outperforms stronger monolithic baselines across five agentic tasks, including mathematical reasoning, financial research, scientific research, accelerator design, and distributed-system optimization. We further provide theoretical insights into how economic dynamics shape agent behaviors, linking local incentives to long-term global performance. Our results suggest a new path to multi-agent intelligence: rather than engineering coordination, we can design decentralized incentive structures under which it automatically emerges.

1 Introduction

The paper asks how bounded agents can collectively solve complex tasks without centralized coordination. It proposes Economy of Minds, where economic incentives drive agents to self-orchestrate, self-evolve, and achieve strong results across five agentic tasks.

  • 1 Introduction: Bounded agents with partial observations, distinct priors, and limited computation cannot reliably solve complex tasks individually.The introduction frames collective problem-solving as necessary when tasks exceed individual capabilities.
  • 1 Introduction: Centralized orchestration creates a coordination bottleneck and single point of failure, while making learning and adaptation increasingly difficult.The passage identifies information and decision-making flow through one orchestrator as the source of these limitations.
  • 1 Introduction: Economy of Minds uses auctions, peer-to-peer payments, and outcome-based wealth to make agents self-orchestrate and self-evolve.Successful contributors accumulate wealth and survive, while ineffective agents are replaced through the system’s evolutionary dynamics.
  • 1 Introduction: 57.0% mathematical reasoning, 60.0% financial research, and 20.0% scientific research accuracy are achieved by EOM from respective starting points of 15.9%, 45.0%, and 5.0%.On accelerator design, EOM reaches an average energy-delay product of 39.3 versus 80.2 for a strong domain-specific method; distributed-system optimization costs 657 versus 930 for the baseline.
  • 1 Introduction: Economic organization enables coordination, specialization, and cooperation to emerge without explicitly designed coordination mechanisms.The authors present this as a general and scalable foundation for decentralized multi-agent intelligence built from evolving agent societies.

2 An Economy of Language Agents

The economy coordinates language agents through decentralized auctions and backward wealth transfers during planning, then evolves the population across episodes through wealth-based exploitation and exploration without centralized supervision.

  • System design: The system couples within-episode planning, which selects actions and assigns credits, with across-episode adaptation, which evolves the agent population.Each agent acts locally from its triggering condition and policy, while economic interactions generate global coordination.
  • Auction-based planning: At each step, eligible agents auction for control, and the highest-bidding agent wins, acts, and advances the environment.If no agent is eligible, the episode terminates or defaults to a null transition, depending on the environment.
  • Economic credit assignment: Bucket-brigade payments create decentralized credit assignment by rewarding actions that enable productive downstream states and penalizing unproductive continuations.The winner pays its bid to the previous winner while collecting environmental reward; the first winner pays the house.
  • Population adaptation: Economic selection preserves wealthy agents, mutates successful prompts, removes bankrupt agents, and injects replacements to reuse effective behaviors while exploring new ones.Agents can lose wealth through unhelpful actions or inactivity, while wealthy agents are selected as parents for mutation.
  • Population adaptation: Evolution is governed entirely by wealth accumulation and loss, enabling self-adaptation without centralized supervision or explicit global performance labeling.The resulting system combines exploitation, which sharpens useful behaviors, with exploration, which injects novelty and prevents premature convergence.

3 Experiments

Across five domains, EOM turns role-specialized, access-limited agents into collective systems that match or outperform complete-agent baselines. Ablations and adaptation analyses indicate that economic selection, value transfer, and population evolution—not agent multiplicity alone—drive these gains and support transfer, specialization, and reusable structure.

  • Experimental setup: EOM is evaluated on mathematical reasoning, financial research, scientific research, accelerator design, and distributed-system optimization using partial agents with restricted capabilities.Partiality includes restricted action spaces, limited tools, short generation budgets, specialized roles, or partial environmental observations.
  • Cross-domain performance: On MATH, EOM improves Llama-3.1-8B agents from 15.9% to 57.0% and Gemma-2-9B agents from 4.2% to 45.1%, exceeding complete-agent baselines of 51.9% and 44.3%.Each population agent is role-specialized and restricted to short outputs.
  • Cross-domain performance: On Finance-Agent-Bench, EOM rises from 45.0% to 60.0% after 30 tasks, while FrontierScience-Research reaches 8.5% mean and 20.0% best-run accuracy.These results outperform the cited Finance-Agent-Bench baselines and exceed GEA’s 1.8% mean and 5.0% best-run accuracy on FrontierScience-Research.
  • Economic dynamics: MATH ablations reduce performance when rent or rewards are perturbed, lowering mean accuracy from 43.9 to 39.0–41.8 and best-run accuracy from 57.0 to 44.0–47.0.The results indicate dependence on the balance among reward inflow, rent pressure, and agent survival.
  • Economic dynamics: On Cloudcast, EOM reaches a best cost of 673 versus 999 for best-of-N multi-agent sampling, showing that repeated sampling alone is insufficient.The best-of-N baseline uses multiple agents but does not evolve them through market selection.
  • Adaptation and robustness: Economic selection preserves useful lineages, removes ineffective agents, sharpens prompts and interaction topology, and enables transfer, specialization, and reusable structure.EOM achieves a 2.2× geometric-mean EDP gain over DOSA across 24 ResNet-50 kernels, while easy-to-hard MATH training finishes around 57% versus about 47% for reversed ordering.

4 Related Work

Related work studies self-evolving agents that improve through interaction, adaptation, accumulated experience, and local feedback rather than fixed centralized control. In LLM populations, this includes co-evolution, iterative refinement, experience sharing, and reward-driven emergence of structured behavior.

  • Self-Evolving Agents: Self-evolving agents improve through interaction, adaptation, and accumulated experience instead of fixed centralized control.
  • Self-Evolving Agents: LLM research explores co-evolution, iterative refinement, and experience sharing across agent populations.
  • Self-Evolving Agents: Reward-driven self-organization shows that structured behaviors can emerge from local feedback alone.

5 Limitations, Conclusions, and Future Work … C.2 Outcome-only reward is sufficient for a strong population

The paper presents decentralized economic interactions as a way to produce adaptive collective intelligence, while identifying frozen-backbone prompt adaptation as a limitation. Its theory and algorithms explain market selection, outcome-only reward sufficiency for strong populations, and the training and evaluation procedures underlying the framework.

  • 5 Limitations, Conclusions, and Future Work: Decentralized auctions, transactions, and wealth-based selection induce specialization and coordination without centralized control across diverse domains.The authors frame incentive design as an alternative to designing individual agents.
  • 5 Limitations, Conclusions, and Future Work: Prompt-space adaptation with a frozen backbone may restrict capability growth when tasks require new skills or representations.Proposed extensions include parameter-space training, hybrid adaptation, and multimodal or embodied backbones.
  • A Pseudo Code: The training loop repeatedly plans through auctions, credits the final actor with environmental reward, removes bankrupt agents, rolls back survivors, and replenishes the population through adaptation.Periodic rent and bankruptcy-triggered or periodic births regulate population evolution.
  • A Pseudo Code: Evaluation freezes agent wealth, bids, and status, then runs thread-local population snapshots without payments, rewards, rent, births, or shared mutation.The evaluation loop records task trajectories, termination, and environment scores in parallel.
  • B Extended Related Works: Related work studies collective coordination through specialization, dynamic roles, reinforcement learning, evolutionary adaptation, debate, verification, and decentralized routing.These approaches generally seek coordination through local interactions rather than centralized planning.
  • B Extended Related Works: The framework draws on economic and distributed-intelligence ideas in which coordinated outcomes emerge from decentralized actors and interactions among many simple agents.Relevant influences include Hayek’s spontaneous order, Minsky’s Society of Mind, bucket-brigade credit assignment, and Hayek Machines.
  • C Theoretical Motivations: Market dynamics select useful specialists because underpriced agents can accumulate wealth and persist, while overpriced agents lose wealth and become insolvent.Bankruptcy and replacement push the surviving bid frontier toward the value of the best available specialist.
  • C.2 Outcome-only reward is sufficient for a strong population: Outcome-only reward suffices when the auction population reliably selects an approximately best eligible specialist at each reachable history.Process rewards may shape training, but are unnecessary for correctness once market selection is near-optimal under the outcome objective.

C.3 Regret to an oracle coordinator · C.4 Bucket-brigade payments recover structured credit assignment · C.5 Malicious agents and collusion

The paper shows that calibrated decentralized auctions can approach an oracle coordinator, while bucket-brigade payments recover structured local credit assignment. Bankruptcy removes value-destroying agents, but collusion shifts selection from individuals to coalitions that may persist when externally profitable.

  • C.3 Regret to an oracle coordinator: As bids approximate specialist action-values, the decentralized auction tracks a centralized oracle coordinator with vanishing regret.The comparison is explicitly against an oracle that directly selects the best eligible specialist at each history.
  • C.3 Regret to an oracle coordinator: Theorem 3 bounds episode regret by r_e ≤ 2β_e P H^-1 when bid overestimation is bounded by a nonincreasing sequence β_e.The bound applies for every episode, reachable history, and eligible agent under the theorem’s stated assumption.
  • C.4 Bucket-brigade payments recover structured credit assignment: Bucket-brigade payments assign credit to earlier agents through the next winner’s payment, which can represent a Bellman continuation value under continuation pricing.Under the stated condition, the winner’s wealth update uses a Bellman-like one-step target.
  • C.4 Bucket-brigade payments recover structured credit assignment: In acyclic workflows, continuation-priced payments coincide, up to a downstream baseline, with each agent’s ordered marginal contribution to final outcome value.This result assumes subtasks execute in topological order and the next winner prices the remaining feasible subgraph.
  • C.4 Bucket-brigade payments recover structured credit assignment: Bucket-brigade transfers propagate downstream value backward through executed agent chains as a local credit-assignment signal.The mechanism therefore does more than redistribute wealth.
  • C.5 Malicious agents and collusion: Bankruptcy removes malicious agents in finite time when aggressive bids exceed expected resale-adjusted payoff and produce negative wealth drift.The rule discourages bids that do not create commensurate value.
  • C.5 Malicious agents and collusion: For colluding agents, selection operates on coalition wealth because internal transfers cancel, and a cartel can survive only with nonnegative aggregate external wealth drift.If every possible surviving cartel is externally loss-making, bankruptcy eliminates collusion; otherwise a profitable cartel may persist or form a monopoly.

C.6 Proofs for Section C.1 · C.7 Proofs for Section C.2

The appendices prove survival and valuation bounds for recurrent auction winners, and relate auction outcomes to future suboptimality through an expectation argument. Together, these proofs establish the theoretical claims in Sections C.1 and C.2.

  • C.6 Proofs for Section C.1: Under stationarity and bounded payoffs, an underpriced agent survives forever after its first win with positive probability.The proof models resale-adjusted payoffs as bounded i.i.d. increments and uses a nonnegative supermartingale.
  • C.6 Proofs for Section C.1: The bankruptcy argument defines the stopping time over wins and bounds the probability of bankruptcy using wealth before the first win.On the bankruptcy event through n wins, the proof applies the supermartingale construction to control survival.
  • C.6 Proofs for Section C.1: Every forever-surviving recurrent winner must have a bid no greater than its value at the context.Otherwise, the strong law implies wealth converges to negative infinity, forcing eventual removal and contradicting survival.
  • C.6 Proofs for Section C.1: If the recurrent solvent frontier remains below the optimal benchmark by a positive margin, infinitely many injection opportunities create profitable entrants.The lower-bound contradiction assumes β∞(x) < V ⋆(x) − εmax − 2δ and uses recurrent opportunities for injection.
  • C.6 Proofs for Section C.1: Each profitable entrant has a positive probability of surviving forever, so infinitely many independent opportunities imply that one eventually survives almost surely.The entrant’s value exceeds its bid by at least δ, and Lemma 1 supplies survival probability at least qδ > 0.
  • C.7 Proofs for Section C.2: For Section C.2, the theorem begins from a reachable history with a nonempty eligible set and applies the stated assumption.The proof introduces the relevant quantity and evaluates the theorem at each reachable history.
  • C.7 Proofs for Section C.2: After the auction winner acts, the remaining loss is identified as future auction suboptimality.Taking expectation over the initial history then proves the theorem.

C.8 Proofs for Section C.3 · C.9 Proofs for Section C.4

The proofs establish the auction’s near-optimality and regret guarantees by combining bid concentration with auction maximization, while the workflow analysis connects continuation payments to ordered downstream value through telescoping. Together, these arguments justify the theoretical results for Sections C.3 and C.4.

  • C.8 Proofs for Section C.3: Theorem 3 fixes an episode and reachable history, comparing the oracle-selected eligible agent with the auction winner.The proof introduces u^orc_e,t and u^auc_e,t as the central objects of comparison.
  • C.8 Proofs for Section C.3: Bid concentration bounds the relevant agents’ bids, enabling a comparison between the auction winner and the oracle-selected agent.The argument applies bid concentration before and after using the auction’s bid-maximization property.
  • C.8 Proofs for Section C.3: Combining the three inequalities shows that the auction winner is 2β_e-optimal at every reachable history.This is the proof’s key per-episode guarantee.
  • C.8 Proofs for Section C.3: Applying Theorem 2 with ε = 2β_e and summing across episodes yields the cumulative regret bound.The proof then divides by E to obtain the average-regret rate.
  • C.9 Proofs for Section C.4: Proposition 1 begins from its stated assumption and substitutes the resulting identity before taking conditional expectations.The supplied proof text presents this as the initial algebraic step.
  • C.9 Proofs for Section C.4: For an acyclic workflow, every execution induces a topological order, and completing subtask u expands the completed prefix from S to S∪{u}.This structural observation frames the continuation-value argument.
  • C.9 Proofs for Section C.4: Under continuation pricing, the successor’s willingness to pay tracks the expanded feasible subgraph’s continuation value up to a downstream baseline.Thus, payment flows back to the current agent according to the downstream value unlocked by executing u.
  • C.9 Proofs for Section C.4: Summing these ordered margins over the realized topological order produces a telescoping decomposition of the final outcome value.The decomposition holds up to the baseline term.

C.10 Proofs for Section C.5 … D.4 Scientific Research

The appendix proves that malicious agents and externally loss-making collusions are eliminated almost surely under recurrent rent and negative external wealth drift. It also specifies EOM’s partial-agent evaluation framework, baselines, and role-specialized setups for mathematical, financial, and scientific research tasks.

  • C.10 Proofs for Section C.5: Malicious agents are removed in finite time almost surely when they incur a persistent payoff deficit at recurrent eligible contexts and positive rent is charged infinitely often.The proof covers both infinitely many wins, where wealth tends to negative infinity, and finitely many wins, where rent alone eventually causes bankruptcy.
  • C.10 Proofs for Section C.5: Coalitions whose every surviving sub-coalition has strictly negative expected aggregate external wealth drift are eliminated in finite time almost surely.Internal bucket-brigade transfers cancel in coalition wealth, while bounded negative-drift increments drive surviving coalitions below zero through successive bankruptcies.
  • D Additional Experiment Details: EOM is evaluated primarily at inference time against complete-agent, partial-agent, and domain-specific baselines using local H200 computation and commercial model APIs.The implementation uses at most one NVIDIA H200 GPU per experiment and includes official APIs from OpenAI, Google Gemini, and Anthropic Claude.
  • D.1 Baseline Details: The baseline suite tests EOM against complete-agent reasoning, self-improvement, monolithic evolution, non-economic multi-agent interaction, repeated sampling, and accelerator-specific search.REACT and GEA provide complete-agent comparisons; OPENEVOLVE targets Cloudcast, Multi-Agent Debate omits economic mechanisms, best-of-N matches EOM’s episode count, and DOSA targets accelerator design.
  • D.2 Mathematical Reasoning: Mathematical reasoning uses an easy-to-hard MATH training stream with role-specialized planner, executor, and verifier agents constrained to 64–256-token outputs.Experiments use weak Llama-3.1-8B and Gemma-2-9B backbones, compared with corresponding complete agents having full output and action capability.
  • D.3 Financial Research: Financial research uses four-tool company-filing tasks in which each partial agent has exactly one tool, requiring coordination among specialized search, parsing, retrieval, and submission roles.The evaluation selects 30 training tasks and 20 test tasks, comparing EOM with Multi-Agent Debate and complete-agent REACT and GEA baselines.
  • D.4 Scientific Research: Scientific research evaluates FrontierScience with literature, planner, executor, and verifier agents that divide background knowledge, decomposition, sub-question solving, and validation.The Research subset uses a 40/20 train–test split, Gemini-3-Flash, comparison with GEA, and mean- and best-run accuracy across evolution checkpoints.

D.5 Accelerator Design · D.6 Distributed-System Optimization · D.7 Evaluation Protocol

The evaluation covers accelerator mapping and distributed-system code optimization, using role-specialized partial agents and auction-based execution. Test-time results are measured with the trained population frozen, separating learned coordination from continued adaptation.

  • D.5 Accelerator Design: EOM searches hardware mappings for 24 ResNet-50 convolution kernels in GEMMINI, minimizing energy-delay product (EDP).Lower EDP indicates a better accelerator design.
  • D.5 Accelerator Design: Accelerator exploration uses Historians, Planners, and Executors to divide memory, architectural, and fine-grained evaluation responsibilities.Historians maintain promising or failed design directions, while Planners propose search directions and Executors perform local evaluations.
  • D.5 Accelerator Design: EOM is compared with DOSA and a complete REACT agent on average EDP across the 24 kernels, with lower values indicating better designs.Both LLM-based methods use the Gemma-4-31B backbone; Figure 5 analyzes wealth dynamics and Figure 7 reports per-kernel EDP improvements.
  • D.6 Distributed-System Optimization: Cloudcast minimizes total data-transfer cost through iterative code optimization, starting each episode from the last successfully verified program and receiving reward at the end.The setup resembles test-time reinforcement learning because improvement attempts begin from a parent solution and use reward signals.
  • D.6 Distributed-System Optimization: Cloudcast uses specialized Planner, Reader, Implementer, Builder, Evaluator, and Finalizer agents for staged coding, testing, verification, and submission.Each partial agent is responsible for one stage of the coding and verification workflow.
  • D.6 Distributed-System Optimization: EOM runs Cloudcast for 30 episodes versus OPENEVOLVE’s standard 300 iterations, alongside a best-of-N multi-agent baseline.All methods use GPT-5-mini, and OPENEVOLVE’s setup takes longer in wall-clock time.
  • D.7 Evaluation Protocol: During training, agents auction for actions within episodes, exchange payments, and adapt across episodes through rent, bankruptcies, exploration, and exploitation.Agents with negative wealth are removed, while new agents are injected through exploration and exploitation.
  • D.7 Evaluation Protocol: During evaluation, the trained population is frozen, and auctions select actions without payments, rewards, rent, births, bankruptcies, or further adaptation.Each test task uses a thread-local copy, so results measure the learned society at a fixed checkpoint.

E Prompt and Topology Evolution in Scientific Research

In FrontierScience-Research, economic selection evolves an EXECUTER prompt from generic algebraic execution into a reusable, self-auditing scientific reasoning routine. As this competence transfers across domains, auction-selected topologies contract or expand conditionally, while population-level selection allocates future mutations toward agents that repeatedly create downstream value.

  • E.3 From Local Reasoning Evolution to Topology Evolution: As EXECUTER internalizes auditing and principle identification, separate VERIFIER and LITERATURE turns become less necessary in some states, changing the wakeup landscape.The marginal value of those roles decreases because dimensional checks, equation-vs-DOF counting, constraint verification, falsification, and principle naming occur internally.
  • E Prompt and Topology Evolution in Scientific Research: Across 40 training episodes, descendants of one evolving EXECUTER carry nine of eleven successful episodes, providing a traceable view of prompt-level evolution.Each child inherits its parent’s trainable prompt and receives only a small mutation.
  • E.1 Prompt-Level Evolution: From Generic Execution to Reusable Scientific Reasoning: The prompt becomes operational rather than merely longer, adding reusable procedures for explicit decomposition, constraints, well-posedness, and falsification.The final routine identifies the governing principle, checks symmetries, verifies well-posedness, expands explicitly, enforces constraints, and substitutes results back into governing equations.
  • E.1 Prompt-Level Evolution: From Generic Execution to Reusable Scientific Reasoning: Coordinate-based, component-wise execution exposes scalar variables and equations, making intermediate reasoning easier for VERIFIER agents to audit.This mutation was distilled from a CMB trajectory where separate scalar relations proved more reliable than manipulating one high-level covariance object.
  • E.2 Cross-Domain Transfer of the Evolved Reasoning Routine: The evolved routine transfers from physics and chemistry tasks to spectroscopy, biology, and synthesis, including a 195Pt NMR task unrelated to its shaping domains.In episode 32, it applies governing-principle identification, equation-versus-unknown counting, symmetry checks, scalar expansion, and substitute-back validation.
  • E.4 Early Regime: Explicit Multi-Role Auditing: Early successful trajectories rely on explicit multi-role auditing: episode 11 scores 0.75 in ten auction steps using all five roles.Its workflow emerges from local wakeup judgments, novice bidding, same-role blocking, and the terminal answer restriction rather than global workflow design.
  • E.6 Training-Wide Topology Trend: Later trajectories often contract to compact 3–4 step paths, but longer literature and verification chains remain when residual uncertainty is high.This topology change reflects concentrated competence among load-bearing agents, not a smaller available population.
  • E.7 Population-Level Selection: Economic feedback deepens EXECUTER descendants and broadens ANSWER families, while comparatively shallow LITERATURE and PLANNER families receive less evolutionary expansion.Agents creating downstream value accumulate wealth, survive longer, and become more likely to seed future mutations, allocating future evolutionary capacity by repeated success and failure.

F Prompt and Topology Evolution in CloudCast … G.2 Evolved Generalist Prompt

Across CloudCast and Finance-Agent-Bench, economic evolution adapts both action topology and prompts to local uncertainty and costly decisions. In the generalist case, evolution improves rigor through broad compliance rules, but does not create a sharply specialized competitor.

  • F Prompt and Topology Evolution in CloudCast: CloudCast evolves an action discipline for avoiding expensive actions, with topology adapting between short improvement paths and longer uncertain search or repair phases.The learned object is when to act, not a new routing heuristic.
  • F.1 Auction-Derived Topology: Six roles compete through wakeup judges and fixed bids, with the highest bidder acting and consecutive same-role turns allowed to emerge.The roles are READER, PLANNER, IMPLEMENTER, BUILDER, EVALUATOR, and FINALIZER.
  • F.1 Auction-Derived Topology: Topology length tracks residual uncertainty rather than training time: Episode 9 collapses to four steps, whereas Episode 15 uses the full step budget.Short paths reflect one-edit improvements; long paths reflect multi-edit search or regression repair.
  • F.2 Prompt Evolution: Checks Before Costly Actions: Prompt mutations teach EVALUATOR, BUILDER, IMPLEMENTER, and FINALIZER to perform cheap structural checks before expensive actions.These checks cover evaluation markers, build symbols, intent-and-verify editing, and final invariants.
  • G Why the Generalist Does Not Monopolize: The economy’s generalist experiment tests whether full access to edgar_search, web_search, parse_html_page, and retrieve_information causes market monopolization.The generalist is evaluated alongside ordinary specialized agents in Finance-Agent-Bench.
  • G.1 Initial Generalist Prompt: The initial generalist prompt enforces tool-call correctness and search-first sequencing but lacks domain-specific decomposition, evidence hierarchy, period alignment, and financial-error safeguards.It functions as a broad operator’s manual rather than a financial reasoning procedure.
  • G.2 Evolved Generalist Prompt: Evolution adds broad requirements for decomposition, interval coverage, consolidated totals, raw-output tracing, accounting consistency, discrepancy checks, and trend sanity checks.These rules make the generalist safer and more rigorous, but its prompt remains a broad compliance layer rather than a specialist heuristic.

G.3 Comparison to Specialized Agent Evolution · G.4 Generalist Failure as Evidence for Specialization

Specialized agents evolve toward sharper, domain-specific decision rules, while the full-tool generalist accumulates broad obligations and generic caution. Because auctions reward locally high-value behavior, specialization wins control where over-generality dilutes local advantage.

  • G.3 Comparison to Specialized Agent Evolution: Specialized EDGAR and TAVILY agents remain narrow after evolution but become more exacting about failure modes within their respective tool domains.EDGAR focuses on filings, while TAVILY focuses on targeted web search.
  • G.3 Comparison to Specialized Agent Evolution: EDGAR evolves toward filing-specific correctness by checking entities, filing types, fiscal periods, qualifiers, dates, projections, and aggregate-versus-segment values.Its prompt does not attempt to become a universal analyst.
  • G.3 Comparison to Specialized Agent Evolution: TAVILY evolves toward epistemic discipline by identifying knowledge gaps, tracing claims to primary evidence, matching periods and metrics, cross-referencing sources, and recomputing arithmetic.The resulting specialization remains within one tool domain rather than moving toward generality.
  • G.3 Comparison to Specialized Agent Evolution: The full-tool generalist must absorb heterogeneous concerns simultaneously, so its prompt accumulates global obligations instead of repeatedly repairing one local competence.This broad responsibility dilutes its local edge under auctions that reward agents with high value in particular states.
  • G.4 Generalist Failure as Evidence for Specialization: Prompt evolution works best when repeated feedback compresses into a small number of role-specific decision rules targeting recurring mistakes.Narrow interfaces and evidentiary responsibilities make mutations concrete, sharper, more falsifiable, and more economically competitive.
  • G.4 Generalist Failure as Evidence for Specialization: Generalist mutations add generic caution across heterogeneous subtasks, making prompts longer, broader, and procedurally safer without strengthening one specific inference pattern.The generalist therefore fails to gain the same specialization advantage.
  • G.4 Generalist Failure as Evidence for Specialization: Specialization wins control in this market because over-generality does not produce economic dominance or monopolization.The generalist does not monopolize when its responsibilities span too many heterogeneous subtasks.
Loading 2606.02859v1…