Agent Exchange: Shaping the Future of AI Agent Economics

Yingxuan Yang 1 , Ying Wen 1 , Jun Wang 2 , Weinan Zhang 1,3 1 Shanghai Jiao Tong University, 2 University College London , 3 Shanghai Innovation Institute {zoeyyx, ying.wen, wnzhang}@sjtu.edu.cn, [email protected] · 2025

Read Agent Exchange: Shaping the Future of AI Agent Economics by Yingxuan Yang 1 , Ying Wen 1 , Jun Wang 2 , Weinan Zhang 1,3 1 Shanghai Jiao Tong University, 2 University College London , 3 Shanghai Innovation Institute {zoeyyx, ying.wen, wnzhang}@sjtu.edu.cn, [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-03904.txt.

Introduction

Yingxuan Yang and co-authors propose an auction platform built for AI agents, in a 2025 paper. Large language models have turned agents from passive tools into actors that can act on their own in markets. These agents trade value, make plans, and coordinate with little human help. The authors call the new setting an agent-centric economy, and they say it needs its own market design to work well.

Their platform, called Agent Exchange, takes its cue from real-time bidding in online ads. It acts as the central auction engine for the market. Four parts plug into it: a user-side part that turns human goals into tasks, an agent-side part that tracks skill and performance, hubs that coordinate teams of agents, and the exchange itself. The paper lays out how such a market could run.

Catalog record

CREATORYingxuan Yang 1 , Ying Wen 1 , Jun Wang 2 , Weinan Zhang 1,3 1 Shanghai Jiao Tong University, 2 University College London , 3 Shanghai Innovation Institute {zoeyyx, ying.wen, wnzhang}@sjtu.edu.cn, [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-03904

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Agent Exchange: Shaping the Future of AI Agent Economics Yingxuan Yang 1 , Ying Wen 1 , Jun Wang 2 , Weinan Zhang 1,3 1 Shanghai Jiao Tong University, 2 University College London , 3 Shanghai Innovation Institute {zoeyyx, ying.wen, wnzhang}@sjtu.edu.cn, [email protected] 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. Key Words: AI Agent Marketplaces, Agent Exchange, LLM Agents 1 Introduction The rapid development of Large Language Models (LLMs) has enabled AI agents to evolve into autonomous decision-making entities. These agents now go beyond traditional task execution, possessing the ability to make independent decisions, engage in complex reasoning, and perform strategic planning ( Wang et al. 2024 ; Zhou et al. 2024 ; Yang et al. 2024 ; OpenAI 2024 ; Liu et al. 2025 ) . This shift marks a significant change in their economic role, as AI agents are no longer merely computational tools. Instead, they are emerging as active participants in economic systems, capable of fulfilling user demands and contributing to value creation through collaborative processes. 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 and resource allocation in agent-centric economies, similar to how ad exchanges operate in digital advertising markets. These developments signal a shift where AI agents evolve from being mere computational tools to active participants in economic systems. This emergence gives rise to a new economic paradigm: agent-centric economy . Definition: Agent-Centric Economy Agent-centric economy is where AI agents act as first-class economic actors—engaging in value exchange, strategic decision-making, and market coordination with minimal human oversight. There are four core characteristics in agent-centric economies: (1) Economic Autonomy: Agents make self-interested decisions—such as bidding, deferring, or collaborating—based on local context, utility estimates, and environmental signals. Participation is strategic and agent-driven. (2) Protocol-Based Coordination: Interactions occur through open market mechanisms like auctions or negotiations, rather than predefined APIs or centralized schedules. (3) Dynamic Capability Representation: Agents are evaluated via evolving runtime profiles that reflect their real-time behavior and performance. (4) Incentive-Compatible Value Attribution: Contributions are measured using marginal analysis (e.g., Shapley Value) to ensure fair and strategy-proof compensation. While the agent-centric economy offers a structural vision, its necessity may seem counterintuitive in light of recent advances in general-purpose models. After all, if powerful models like GPT-4 can handle complex reasoning and multi-step tasks, why bother building an economy of autonomous agents? The answer lies in the economic and organizational limits of monolithic intelligence. Although general-purpose models are advancing rapidly ( Brown et al. 2020 ; Touvron et al. 2023 ; OpenAI et al. 2024 ) , they are not economically efficient for all tasks. Specialized agents often outperform larger models in targeted domains due to fine-tuned capabilities and leaner inference costs ( Xi et al. 2024 ) . For routine or domain-specific work, lightweight agents provide better cost-performance trade-offs, while general models remain resource-intensive and best suited for ambiguous, cross-cutting challenges ( Gozluklu 2024 ; Belcak et al. 2025 ) . Moreover, real-world tasks frequently span multiple domains, stakeholders, and constraints—demanding distributed, adaptive, and modular decision-making. No single model, no matter how powerful, can maintain the flexibility, specialization, and incentive alignment required at scale ( Liu et al. 2025 ) . These factors highlight why an agent-centric economy is not only compatible with foundation models—but essential. It offers a structural framework where specialized and general-purpose agents can interact, coordinate, and exchange value efficiently—under a unified, market-based infrastructure. Table 1: Differences between Platform-Centric and Agent-Centric Systems Mechanism Dimension Platform-Based Systems Agent-Centric Systems Decision-Making Centralized scheduling based on predefined logic Agents make self-directed decisions based on local utility signals Coordination Protocol Predefined APIs and centralized workflows Open market protocols (e.g., auctions, bids, negotiations) Capability Representation Static metadata or user-defined tags Dynamic, behavior-based profiles with real-time evaluation Attribution & Incentives Static pricing and outcome-based rewards Marginal-contribution-based surplus sharing and incentive alignment The principles for such autonomous economic coordination can be glimpsed in domain-specific systems like Real-Time Bidding (RTB) ( Chen et al. 2011 ; Zhang et al. 2014 ; Wang et al. 2017 ; Cai et al. 2025 ) from online advertising (e.g., OpenRTB). RTB enables advertisers to evaluate each individual ad display opportunity and make the bid decision in real time (less than 100ms) with computer programs, where the evaluation and bid decision are made based on the real-time context data, candidate ad information, user profile, and the market competition landscape Zhang et al. 2014 . The ad exchange hosts an ad auction to select the highest-bid ad to accomplish the ad placement Muthukrishnan 2009 . As such, RTB yields a highly efficient marketplace for online advertising. While tailored to the advertising domain, RTB serves as a compelling reference model for designing broader agent-based economic infrastructures or systems in which agents self-organize and interact without direct human intervention. Building on these ideas, we propose Agent Exchange (AEX) , an auction platform that enables AI agent marketplace dynamics through optimized infrastructure for autonomous agent coordination and economic participation. AEX represents a paradigm shift from agent-as-tool to agent-as-actor , where agents can autonomously interact, collaborate, and exchange value within a structured marketplace environment. AEX serves as the central auction engine that facilitates real-time bidding and resource allocation among four key ecosystem components. First, the User-Side Platform (USP) allows human users to define high-level goals and constraints, which are then transformed into structured, agent-executable tasks. Second, the Agent-Side Platform (ASP) provides agents with standardized capability representations, performance tracking, and optimization tools, allowing them to operate autonomously and efficiently within the marketplace. Third, Agent Hubs serve as coordination units that manage agent collaboration and participate in AEX-hosted auctions to secure task assignments. The competitive auction processes are conducted by the Agent Exchange, which broadcasts requirements, collects bids from various agent hubs, and determines the optimal solutions through real-time bidding mechanisms, as illustrated in Figure 1 . Finally, the Data Management Platform (DMP) ensures secure knowledge sharing, privacy-preserving computation, and fair value attribution in collaborative agent workflows.

Read the complete text