Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

Agent Exchange: Shaping the Future of AI Agent Economics

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proposes Agent Exchange (AEX), an auction platform where autonomous AI agents bid for tasks, form teams, and receive credit by marginal contribution, and argues it can be the central coordination mechanism of an agent-centric…

desk verdict A coherent architectural vision for agent economies, but the simulation validates the auction with the auction's own objective, so the empirical claims are largely circular and the real value is as a framing paper. read the letter →

arxiv 2507.03904 v1 pith:4K23SD4O submitted 2025-07-05 cs.AI cs.MA

Yingxuan Yang , Ying Wen , Jun Wang , Weinan Zhang This is my paper · ORCID
classification cs.AIcs.MA
keywords AIagentmarketplacesExchangeLLMagentsagent-centriceconomyreal-timebiddingmulti-attributeauctionsShapleyvalueadaptivemechanismselection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper is trying to establish that the next stage of AI is an agent-centric economy, in which autonomous agents act as first-class economic participants, and that a dedicated auction platform'Agent Exchange (AEX)' should be the central coordination mechanism of that economy. The claim matters because it shifts the bottleneck from raw model capability to market infrastructure: specialized agents and general-purpose models would meet through bidding, price discovery, and fair credit rather than through predefined APIs. AEX is designed as an exchange inspired by real-time bidding in online advertising, connecting a user-side platform that turns human goals into tasks, an agent-side platform that represents and verifies capabilities, agent hubs that bid and coordinate, and a data platform for secure knowledge sharing and value attribution. Its key operating rule is adaptive mechanism selection: auction when competition is sufficient, assign directly when it is not. A one-shot simulation comparing five allocation strategies is presented as proof-of-concept that the enhanced auction delivers the most balanced quality, cost, and robustness profile.

What carries the argument

The load-bearing machinery is AEX's two-stage adaptive allocation engine. At each stage it chooses between a competitive auction and direct assignment based on market liquidity; when auctioning, it scores bids with a multi-attribute composite score $S_{ij}=\sum_k w_k f_k(a_i,s_j)\,\varphi(w_j,\alpha_j)$ that folds capability match, quality, cost, and time into one number, and charges each winner a generalized second-price payment' the second-highest composite score plus a small increment $\epsilon$' which the paper asserts encourages truthful bidding. For team formation it maximizes expected net value $\max \sum_{S\subseteq A}[p_{\text{success}}(S)v(S)-c(S)]x_S$ over available agents, and for reward splitting it uses the Shapley value $\phi_i=\sum_{S\subseteq A\setminus\{i\}}\frac{|S|!(|A|-|S|-1)!}{|A|!}[v(S\cup\{i\})-v(S)]$, which credits each agent by its average marginal contribution across all possible orderings. The combination of adaptive switching, composite scoring, and marginal-contribution attribution is what carries the paper's claim that AEX can deliver efficient allocation and fair compensation without a central planner.

What would settle it

Re-run the same one-shot allocation simulation with a share of agents that inflate their capability scores (for example, multiplying $C_{ic}$ by 1.2–1.3 for some capabilities) and with noisy information about peer capabilities; if the enhanced auction's allocative efficiency and Shapley-based payouts drop to the level of greedy or random allocation, the honest-reporting and perfect-information assumptions are doing the load-bearing work and the central claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that AI agents have crossed from being computational tools to autonomous economic actors, and that this transition requires an open marketplace with a central auction engine rather than centralized scheduling. The proposed engine, Agent Exchange (AEX), treats tasks as tradable items: the User-Side Platform converts human requests into structured task specifications, the Agent-Side Platform maintains standardized capability profiles with verification, Agent Hubs form teams and bid in AEX-hosted auctions, and the Data Management Platform supports secure knowledge sharing and attribution. AEX's specific innovation is adaptive mechanism selection'it runs multi-attribute generalized second-price auctions when hub-level or agent-level liquidity is sufficient, and switches to direct capability-based assignment when competition is scarce. In the paper's simulation, the enhanced auction mechanism produces a quality score of $0.934 \pm 0.062$, cost efficiency of $0.0006$, and robustness of $0.089$, and the authors describe it as the most balanced of the five compared strategies. The empirical claim is deliberately limited to a proof-of-concept under static capabilities, perfect information, and honest reporting.

Load-bearing premise

The load-bearing premise is that agents bid truthfully and report their capabilities honestly, with complete information about tasks and peers; the paper explicitly assumes this, and if agents shade bids or inflate claims, the claimed efficiency and fairness do not follow.

Editorial extensions

If this is right

  • Auction infrastructure can be reused for agent coordination: standardized messaging, millisecond-scale bidding, and second-price payments transfer directly from display advertising to task allocation.
  • Market liquidity becomes a design parameter, since the exchange can degrade gracefully from full competition to direct assignment instead of failing when few agents participate.
  • Capability representation and verification become necessary market infrastructure, because agents can only be compared and fairly compensated if their skills are described in a standard, auditable form.
  • Collaborative surplus can be distributed by marginal contribution, which gives agents a formal reason to join teams rather than work alone.
  • The four auction/assignment configurations map onto real market structures such as competitive platforms, outsourcing, multi-tier supply chains, and long-term partnerships, so the system can be adopted incrementally.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural follow-up the paper does not pursue is to prove whether the $\epsilon$-increment generalized second-price rule remains truthful when bids are multi-attribute composite scores rather than single prices.
  • The adaptive switching rule suggests a concrete experiment not run in the paper: vary the liquidity threshold $N$ and measure the market density at which auction overhead outweighs the gains from price discovery.
  • If the honest-reporting assumption is relaxed, Shapley-based attribution becomes open to manipulation through inflated success claims; a stress test with misreported capabilities would quantify how much fairness degrades.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Agent Exchange (AEX), an auction-based platform intended to serve as the central coordination mechanism of an 'agent-centric economy' in which LLM-based agents act as autonomous economic participants. The architecture comprises four components (USP, ASP, Agent Hubs, DMP) and centers on an adaptive auction engine that switches between competitive bidding and direct assignment depending on market liquidity. The paper contributes a system design, a multi-attribute scoring rule for allocation, a Shapley-value-based attribution scheme, and a preliminary simulation comparing five allocation strategies. The empirical study is explicitly a one-shot proof-of-concept; its main claim is that the Enhanced Multi-Attribute Auction provides the most balanced performance profile across quality, cost efficiency, and robustness.

Significance. The paper addresses a timely and consequential question: how to structure economic interaction among autonomous AI agents. Its RTB-inspired architecture and the explicit decomposition into USP/ASP/Hub/DMP roles provide a useful conceptual vocabulary for agent marketplaces, and the two-stage hub-and-agent auction configuration is a sensible way to frame hierarchical coordination. If the claimed efficiency and fairness properties were rigorously established, the work would be a meaningful step toward practical agent-economy infrastructure. However, the current empirical support is narrow: the simulation tests only one-shot allocation, it does not exercise the adaptive switching mechanism or the two-stage hub auctions that are central to the design, and the outcome model is structurally coupled to the auction's scoring function. These issues currently limit the paper to a systems/vision contribution rather than a validated mechanism design result.

major comments (3)
  1. [§5.1.5 and §5.2.1] The empirical support for the Enhanced Auction is weakened by a structural coupling between the mechanism and the outcome model. In §5.1.5 the auction score S_ij is a weighted sum that includes the capability-matching function f_cap(a_i,s_j), and in §5.2.1 the simulated quality outcome Q_ij = min(P_ij·U(0.9,1.1) + 0.1·f_cap(a_i,s_j), 1.0) adds the same f_cap as a bonus. Because the auction is selecting on f_cap, an allocation that the Enhanced Auction favors will receive a higher Q_ij by construction, so the higher overall performance and the significant pairwise advantage over Random in Table 6 and §5.3 do not by themselves demonstrate an independent efficiency gain. The experiment should either remove the f_cap bonus from Q_ij, define an outcome metric from a separate ground-truth quality model that is not an input to the auction, or report results under both specifications to show that the qualitative conclusions are insensitive to this choice.
  2. [§5 (intro) and §4.3] The simulation validates only a one-shot allocation among five strategies, yet the paper's principal design claim is adaptive mechanism selection: switching between competitive auction and direct assignment depending on market liquidity. Section 5 explicitly restricts the study to a single-round task-to-agent allocation, and none of the experiments exercise the switching rule, the two-stage hub-and-agent configurations in Figure 3, or the hub-level auctions. As a consequence, the central architectural claim that AEX enables efficient, fair coordination through adaptive selection is not tested by the reported proof-of-concept. The paper should either add experiments for the switching rule and the hub configurations, or explicitly reframe the conclusions as describing only the auction-vs-baseline allocation comparison so that the scope of the validation matches the strength of the claims.
  3. [§4.3] The incentive-compatibility claim for the generalized second-price payment rule is not supported. The text in §4.3 states that payment equal to the second-highest composite score plus a small increment epsilon encourages truthful bidding, but no proof or citation is supplied for a multi-attribute scoring rule of the form S_ij = Σ_k w_k f_k(·) φ(·), and the simulation assumes honest reporting and perfect information (§5). Additionally, the adaptive switch between auction and direct assignment creates an incentive for agents to enter or exit competition in ways that the one-shot GSP argument does not cover. The paper should either provide a formal incentive analysis under the composite score, or clearly qualify the claim as a design assumption to be established in future work.
minor comments (5)
  1. [Figure 3] Figure 3 is dense and the small font size makes the auction/direct-assignment labels and the per-agent payment amounts difficult to read; enlarging the figure or splitting it into subpanels would help the reader verify the four configurations.
  2. [§5.2.1] The relation between P_ij and Q_ij is unclear: P_ij is called 'success probability' while Q_ij is called 'quality outcomes for successful executions,' but the reported Quality metric in Table 6 appears to aggregate Q_ij without explaining how failed executions are counted; please clarify the aggregation protocol.
  3. [§5.1.2] The notation C_i is used both for the dynamic capability profile in §3.2 and for the capability strength in the inference equation in §5.1.2; these should be distinguished to avoid ambiguity.
  4. [Table 6] The 'Adaptability' column is reported as High/Medium/Low without a definition or a quantitative basis; please specify how adaptability was determined or omit the column.
  5. [References] The Udinmwen reference is incomplete: the entry ends mid-sentence with '98' and the URL is truncated; the full survey details should be provided.

Circularity Check

1 steps flagged · score 6.0 of 10

The simulated quality outcome includes the same f_cap matching score that the Enhanced Auction optimizes, so the quality-based validation is partially circular.

  1. self definitional [Section 5.1.5 (Allocation Algorithm Implementation) and Section 5.2.1 (Execution Modeling), equations for S_ij and Q_ij.]
    "Enhanced Multi-Attribute Auction ... Sij = ∑_k w_k·f_k(ai,sj)·ϕ(wj,αj) ... fcap(ai,sj) = min(1/|Rj| ∑_{c∈Rj} C_ic, 1.0). ... Quality outcomes for successful executions follow: Qij = min (Pij·U (0.9, 1.1) + 0.1·fcap(ai,sj), 1.0), where ... the capability bonus rewards well-matched assignments."

    The exact capability-matching score f_cap is one of the attributes in the Enhanced Auction's weighted score S_ij, and the same f_cap is added directly to the simulated quality outcome Q_ij with coefficient 0.1. Since the auction ranks by S_ij, it preferentially selects high-f_cap assignments, and those assignments mechanically receive the 0.1·f_cap quality bonus. Thus the quality column in Table 6, which underlies the claim that the Enhanced Auction provides the 'most balanced performance profile,' is not an independent measure of task success: part of its advantage over Random and other baselines is written into the evaluation equation. The paper's own phrase, 'the capability bonus rewards well-matched assignments,' makes the coupling explicit.

full rationale

The main circularity is a structural coupling between the mechanism under test and the evaluation metric: the same f_cap score appears in the Enhanced Auction's scoring function S_ij and in the simulated quality outcome Q_ij as a 0.1·f_cap bonus. Because the auction selects high-f_cap assignments by construction, its measured quality advantage is partially predetermined. I found no other load-bearing circular step: the cited prior work is contextual rather than invoked as a uniqueness theorem or as an unverified foundation for the central claim; the Shapley value is a standard external definition, not repackaged as a prediction; and the stated idealizing assumptions (perfect information, honest reporting, static capabilities) are openly acknowledged limitations rather than hidden inputs to the result. The paper is therefore not wholly circular, but its central empirical validation is weakened by the self-referential quality metric, warranting a score of 6.

Assumptions & free parameters 6 free parameters · 6 assumptions · 2 invented entities

The central claim rests on a set of domain assumptions about agent behavior and capability representation, plus hand-chosen simulation parameters. No external benchmark or real-world data validates the auction mechanisms. The simulation's quality model is partially circular because the auction's own scoring function appears in the outcome equation.

free parameters (6)
  • theta_min = 0.05
    Minimum baseline capability score in the capability inference equation in Section 5.1.2, preventing zero-capability assignments.
  • success probability coefficients = max(0.5, 1.2 - 0.1*alpha_j), upper bound 0.95
    Hand-chosen modulation of base success rate by task complexity in Section 5.2.1, not derived from data.
  • quality noise = U(0.9, 1.1) and N(0, sigma_q^2)
    Synthetic execution variability in Section 5.2.1.
  • capability bonus coefficient = 0.1
    Weight on f_cap in the quality equation in Section 5.2.1, aligning the outcome metric with the auction objective.
  • auction attribute weights w_k = unspecified
    The scoring function S_ij in Section 5.1.5 depends on weights for capability, quality, cost, and time, but the values used in the simulation are not reported.
  • liquidity scenario parameters = high (n_a=9, sigma_q^2=0.05), medium (6, 0.10), low (3, 0.15)
    Market liquidity settings defined in Section 5.1.4.
assumptions (6)
  • domain assumption AI agents will act as first-class economic actors with self-interested utility maximization.
    Foundation of the agent-centric economy premise, stated in the Definition in Section 1.
  • domain assumption Agent capabilities can be represented as dynamic profiles C_i(t) in a standardized latent space.
    Required for capability comparison and matching in Section 3.2, not demonstrated on real agents.
  • standard math A generalized second-price auction with a small increment epsilon encourages truthful bidding.
    Invoked in Section 4.3 from auction theory, but the claim is not proven for multi-attribute composite scores.
  • standard math Shapley value provides fair and strategy-proof compensation.
    Used in Sections 3.3 and 5.2.2; Shapley value is fair by axioms but not generally strategy-proof, so the paper's incentive claims are not supported.
  • domain assumption Real-Time Bidding is a valid reference model for broader agent-based economic infrastructure.
    Basis for the entire AEX design, argued in Section 2.2.
  • ad hoc to paper Market liquidity L = n_a/(sigma_q^2 + 1) is an adequate proxy for auction efficiency.
    Introduced in Section 5.1.4 without validation.
invented entities (2)
  • Agent Exchange (AEX) platform
    purpose: Central auction engine that broadcasts requirements, collects bids, runs auctions, and selects winners in agent marketplaces.
    No implementation, no external benchmarks, and the only support is a synthetic simulation. Provides no falsifiable prediction outside the paper.
  • Agent-centric economy
    purpose: Proposed economic paradigm where AI agents are first-class economic actors exchanging value with minimal human oversight.
    Conceptual framing introduced in Section 1, not an empirical entity and has no measurement outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Agent Exchange: Shaping the Future of AI Agent Economics." pith.science (2026). https://pith.science/paper/4K23SD4O

@misc{pith2026250703904,
  author       = {Pith},
  title        = {Pith review of: Agent Exchange: Shaping the Future of AI Agent Economics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4K23SD4O}},
  note         = {Machine review of arXiv:2507.03904}
}
read the original abstract

The rise of Large Language Models (LLMs) has transformed AI agents from passive computational tools into autonomous economic actors. This shift marks the emergence of the agent-centric economy, in which agents take on active economic roles-exchanging value, making strategic decisions, and coordinating actions with minimal human oversight. To realize this vision, we propose Agent Exchange (AEX), a specialized auction platform designed to support the dynamics of the AI agent marketplace. AEX offers an optimized infrastructure for agent coordination and economic participation. Inspired by Real-Time Bidding (RTB) systems in online advertising, AEX serves as the central auction engine, facilitating interactions among four ecosystem components: the User-Side Platform (USP), which translates human goals into agent-executable tasks; the Agent-Side Platform (ASP), responsible for capability representation, performance tracking, and optimization; Agent Hubs, which coordinate agent teams and participate in AEX-hosted auctions; and the Data Management Platform (DMP), ensuring secure knowledge sharing and fair value attribution. We outline the design principles and system architecture of AEX, laying the groundwork for agent-based economic infrastructure in future AI ecosystems.

Figures

Figures reproduced from arXiv: 2507.03904 by the authors.

Figure 1
Figure 1. Autonomous Economic Decision-Making. This example demonstrates how the Agent Exchange (AEX) facilitates real-time bidding between multiple agent hubs within the AI Agent Marketplace. The exchange broadcasts user requirements, collects competitive bids from different agent hubs, and conducts an auction to select the optimal solution. Hub B wins, illustrating how the auction platform enables efficient price discovery … view at source ↗
Figure 2
Figure 2. AI Agent Marketplace Architecture with Agent Exchange as Core. This figure illustrates the architecture of the AI Agent Marketplace, centered around the Agent Exchange as the core auction platform. The marketplace consists of four key components: User-Side Platform (USP) for request management, Agent Side Platform (ASP) for agent capabilities and bidding, Agent Hub for coordination and resource allocation, and Data … view at source ↗
Figure 4
Figure 4. AEX Workflow in Configuration I. The case study begins with the user submitting a service request, defining constraints like time, budget, and security. In the 2nd stage, the optimal hub is selected through competitive bidding. The 3rd stage involves internal collaboration among agents using a shared knowledge base to complete the task. In the 4th, the final results are delivered, ensuring quality and user acceptanc… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Comprehensive experiment results. The performance of five allocation algorithms across various experimental conditions is illustrated in this figure. It includes comparisons on quality by task complexity, cost efficiency, market liquidity impact, overall method perform…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Blockchain Empowered Trustworthy Agent Networks: Foundations, Taxonomy, and Future Directions

    cs.CR 2026-08 conditional novelty 4.0 of 10

    A survey proposing a five-dimensional taxonomy of trust crises in open AI agent networks and analyzing blockchain's role as a shared trust infrastructure.

  2. BetaWeb: Towards a Blockchain-enabled Trustworthy Agentic Web

    cs.MA 2025-08 unverdicted novelty 4.0 of 10

    BetaWeb promises a blockchain-enabled trustworthy agentic web, but the submitted manuscript body is a different mining-robot paper, leaving the proposal without supporting evidence.

  3. Agentic Web: Weaving the Next Web with AI Agents

    cs.AI 2025-07 conditional novelty 3.0 of 10

    A position paper defines the Agentic Web as the next web era and proposes a three-dimensional conceptual framework for understanding and building it.

Reference graph

Works this paper leans on

31 extracted references · 17 canonical work pages · cited by 3 Pith papers

  1. [2]

    TRAD: Enhancing LLM Agents with Step-Wise Thought Retrieval and Aligned Decision

    URLhttps://arxiv.org/abs/2403.06221. Yingxuan Yang, Qiuying Peng, Jun Wang, Ying Wen, and Weinan Zhang. Llm-based multi-agent systems: Techniques and business perspectives,

    work page Pith review arXiv

  2. [3]

    URLhttps://arxiv.org/abs/2411.14033. OpenAI. Learning to reason with llms.https://openai.com/index/learning-to-reason-with- llms/,

    arXiv

  3. [4]

    Accessed: 2024-04-23. Bang Liu, Xinfeng Li, Jiayi Zhang, Jinlin Wang, Tanjin He, Sirui Hong, Hongzhang Liu, Shaokun Zhang, Kaitao Song, Kunlun Zhu, Yuheng Cheng, Suyuchen Wang, Xiaoqiang Wang, Yuyu Luo, Haibo Jin, Peiyan Zhang, Ollie Liu, Jiaqi Chen, Huan Zhang, Zhaoyang Yu, Haochen Shi, Boyan Li, Dekun Wu, Fengwei Teng, Xiaojun Jia, Jiawei Xu, Jinyu Xian...

    work page 2024

  4. [7]

    URL https://arxiv.org/abs/2303.08774. Zhiheng Xi, Yiwen Ding, Wenxiang Chen, Boyang Hong, Honglin Guo, Junzhe Wang, Dingwen Yang, Chenyang Liao, Xin Guo, Wei He, Songyang Gao, Lu Chen, Rui Zheng, Yicheng Zou, Tao Gui, Qi Zhang, Xipeng Qiu, Xuanjing Huang, Zuxuan Wu, and Yu-Gang Jiang. Agentgym: Evolving large language model-based agents across diverse env...

    arXiv

  5. [8]

    Burak Gozluklu

    URLhttps: //arxiv.org/abs/2406.04151. Burak Gozluklu. How task decomposition and smaller llms can make ai more affordable.Amazon Sci- ence,

    arXiv

  6. [9]

    Ye Chen, Pavel Berkhin, Bo Anderson, and Nikhil R

    URL https://arxiv.org/abs/2506.02153. Ye Chen, Pavel Berkhin, Bo Anderson, and Nikhil R. Devanur. Real-time bidding algorithms for performance-based display ad allocation. In Proceedings of the 17th ACM SIGKDD In- ternational Conference on Knowledge Discovery and Data Mining, KDD ’11, page 1307–1315, New York, NY, USA,

    arXiv

  7. [13]

    ISBN 9798400713316

    Association for Computing Machinery. ISBN 9798400713316. doi:10.1145/3701716.3715259. URL https://doi.org/10.1145/3701716.3715259. S Muthukrishnan. Ad exchanges: Research issues. wine, volume 5929 of lecture notes in computer science,

    arXiv

  8. [15]

    McKinsey Global Survey on AI

    URL https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of- ai-2024. McKinsey Global Survey on AI. 18 Agent Exchange: Shaping the Future of AI Agent Economics Anthony Cardillo. How many companies use ai? (new 2025 data), May

    work page 2024

Show all 31 references
  1. [16]

    Last updated May 28, 2025, Exploding Topics

    URL https: //explodingtopics.com/blog/companies-using-ai . Last updated May 28, 2025, Exploding Topics. PwC US. Pwc’s ai agent survey 2025, May

    2025

  2. [17]

    Accessed: 2025-04-19

    URLhttps://www.anthropic.com/news/model-context- protocol. Accessed: 2025-04-19. Yingxuan Yang, Huacan Chai, Yuanyi Song, Siyuan Qi, Muning Wen, Ning Li, Junwei Liao, Haoyi Hu, Jianghao Lin, Gaowei Chang, Weiwen Liu, Ying Wen, Yong Yu, and Weinan Zhang. A survey of ai agent pr...

    2025 arXiv

  3. [18]

    Accessed: 2025-06-13

    URLhttps://manus.im/. Accessed: 2025-06-13. Salesforce. Agentexchange | salesforce,

    2025

  4. [19]

    Accessed: 2025-06-13

    URL https://www.salesforce.com/agentforce/ agentexchange/. Accessed: 2025-06-13. Shuai Yuan, Jun Wang, and Xiaoxue Zhao. Real-time bidding for online advertising: measurement and analysis. InProceedings of the Seventh International Workshop on Data Mining for Online Advertisin...

    2025

  5. [23]

    Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dongsheng Li, and Deqing Yang

    URLhttps://arxiv.org/abs/2505.10468. Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dongsheng Li, and Deqing Yang. Evoagent: Towards automatic multi-agent generation via evolutionary algorithms,

  6. [24]

    Nikolas Belle, Dakota Barnes, Alfonso Amayuelas, Ivan Bercovich, Xin Eric Wang, and William Wang

    URLhttps://arxiv.org/ abs/2406.14228. Nikolas Belle, Dakota Barnes, Alfonso Amayuelas, Ivan Bercovich, Xin Eric Wang, and William Wang. Agents of change: Self-evolving llm agents for strategic planning,

    arXiv

  7. [25]

    Esther David, Rina Azoulay-Schwartz, and Sarit Kraus

    URL https: //arxiv.org/abs/2506.04651. Esther David, Rina Azoulay-Schwartz, and Sarit Kraus. An english auction protocol for multi- attribute items. InRevised Papers from the Workshop on Agent Mediated Electronic Commerce on Agent-Mediated Electronic Commerce IV, Designing Mec...

  8. [29]

    URL https://arxiv.org/abs/2004.02764. Google. A2a: Agent2agent protocol,

    2004 arXiv

  9. [30]

    Accessed: 2025-04-21

    URL https://github.com/google/A2A. Accessed: 2025-04-21. Smithery. Smithery ai platform,

    2025

  10. [31]

    Accessed: 2025-06-13

    URLhttps://smithery.ai/. Accessed: 2025-06-13. 20

    2025

  11. [1989]

    doi:10.1007/978-1-349-20181-5_25

    ISBN 978-1-349- 20181-5. doi:10.1007/978-1-349-20181-5_25. URL https://doi.org/10.1007/978-1-349-20181- 5_25. Yingxuan Yang, Bo Huang, Siyuan Qi, Chao Feng, Haoyi Hu, Yuxuan Zhu, Jinbo Hu, Haoran Zhao, Ziyi He, Xiao Liu, Zongyu Wang, Lin Qiu, Xuezhi Cao, Xunliang Cai, Yong Yu,...

  12. [2003]

    doi:10.1017/S0269888904000098

    ISSN 0269-8889. doi:10.1017/S0269888904000098. URL https://doi.org/10.1017/ S0269888904000098. Liviu Panait and Sean Luke. Cooperative multi-agent learning: The state of the art. Au- tonomous Agents and Multi-Agent Systems, 11(3):387–434, November

    doi

  13. [2005]

    doi:10.1007/s10458-005-2631-2

    ISSN 1387-2532. doi:10.1007/s10458-005-2631-2. URL https://doi.org/10.1007/s10458-005-2631-2. Ranjan Sapkota, Konstantinos I. Roumeliotis, and Manoj Karkee. Ai agents vs. agentic ai: A concep- tual taxonomy, applications and challenges,

    doi

  14. [2007]

    19 Agent Exchange: Shaping the Future of AI Agent Economics Tim Roughgarden

    doi:10.1109/TASE.2006.885126. 19 Agent Exchange: Shaping the Future of AI Agent Economics Tim Roughgarden. Twenty Lectures on Algorithmic Game Theory. Cambridge University Press,

    2006

  15. [2009]

    The state of ai in early 2024: Gen ai adoption spikes and starts to generate value.McKinsey & Company, May

    Alex Singla, Alexander Sukharevsky, Lareina Yee, Michael Chui, and Bryce Hall. The state of ai in early 2024: Gen ai adoption spikes and starts to generate value.McKinsey & Company, May

    2024

  16. [2011]

    ISBN 9781450308137

    Association for Computing Machinery. ISBN 9781450308137. doi:10.1145/2020408.2020604. URL https://doi.org/10.1145/2020408.2020604. Weinan Zhang, Shuai Yuan, and Jun Wang. Optimal real-time bidding for display advertising. In Proceedings of the 20th ACM SIGKDD International Con...

  17. [2012]

    doi:https://doi.org/10.1016/j.engappai.2011.12.003

    ISSN 0952-1976. doi:https://doi.org/10.1016/j.engappai.2011.12.003. URL https://www.sciencedirect.com/science/article/pii/S0952197611002363. Medet Kanmaz and Elif Surer. Using multi-agent reinforcement learning in auction simulations,

    1976 doi

  18. [2013]

    ISBN 9781450323239

    Association for Computing Machinery. ISBN 9781450323239. doi:10.1145/2501040.2501980. URL https://doi.org/10.1145/2501040.2501980. Iyad Rahwan, Sarvapali D. Ramchurn, Nicholas R. Jennings, Peter Mcburney, Simon Parsons, and Liz Sonenberg. Argumentation-based negotiation. Knowl...

  19. [2014]

    ISBN 9781450329569

    Association for Computing Machinery. ISBN 9781450329569. doi:10.1145/2623330.2623633. URLhttps://doi.org/10.1145/ 2623330.2623633. Jun Wang, Weinan Zhang, and Shuai Yuan. Display advertising with real-time bidding (rtb) and behavioural targeting,

  20. [2017]

    Leng Cai, Junxuan He, Yikai Li, Junjie Liang, Yuanping Lin, Ziming Quan, Yawen Zeng, and Jin Xu

    URLhttps://arxiv.org/abs/1610.03013. Leng Cai, Junxuan He, Yikai Li, Junjie Liang, Yuanping Lin, Ziming Quan, Yawen Zeng, and Jin Xu. Rtbagent: A llm-based agent system for real-time bidding. InCompanion Proceedings of the ACM on Web Conference 2025, WWW ’25, page 104–113, New...

    arXiv

  21. [2020]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971,

    arXiv

  22. [2024]

    doi:10.1007/s11704-024-40231-1

    ISSN 2095-2236. doi:10.1007/s11704-024-40231-1. URLhttp://dx.doi.org/10.1007/s11704-024- 40231-1. Ruiwen Zhou, Yingxuan Yang, Muning Wen, Ying Wen, Wenhao Wang, Chunling Xi, Guoqiang Xu, Yong Yu, and Weinan Zhang. Trad: Enhancing llm agents with step-wise thought retrieval and...

    doi

  23. [2025]

    Tom B Brown, Benjamin Mann, Nick Ryder, et al

    URL https://arxiv.org/abs/2504.01990. Tom B Brown, Benjamin Mann, Nick Ryder, et al. Language models are few-shot learners.Advances in Neural Information Processing Systems, 33:1877–1901,

    1901 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.