AI Agent Economics

Research series · Four working papers 研究专题 · 四篇工作论文

AI Agent Economics: From Delegation Rights to Social Welfare AI Agent 经济学：从委托权到社会福利

As AI agents move beyond advice to use accounts, invoke tools, and complete tasks on behalf of people, economics must revisit several basic questions: who may authorize them to act? How should information and computational resources be organized? Who detects erroneous outcomes and bears their cost? And when execution is hidden, how can we measure its effects on human welfare? 当 AI agent 不再只是提供建议，而是开始代表人使用账户、调用工具并完成任务，经济学需要重新回答几个基本问题：谁有权授权它行动？信息和计算资源应如何组织？错误结果由谁发现并承担代价？当执行过程不可见时，我们又该如何衡量它对人的福利产生了什么影响？

Explore the papers阅读系列论文
01 The question研究问题

Series overview 系列概览

This series studies these questions through four connected areas: delegation rights, the organization of production, execution control, and public policy. Delegation Rights examines control among users, agent providers, and platforms; The Organization of Inference studies how task information and inference budgets shape production; How Do Agent Harnesses Create Value? examines how planning and verification affect success, false passes, and cost; and Welfare-Opaque Income asks how tax policy can assess welfare when the agent's execution rule is hidden from the government.

Using different theoretical and experimental methods, the four papers advance a shared research agenda: to understand how AI agents change rights to act, value creation, and the measurement of outcomes—and to design institutions that are more effective and accountable.

这一论文系列从委托权、生产组织、执行控制和公共政策四个层面研究这些问题。《Delegation Rights》分析用户、agent 提供者与平台之间的控制权；《The Organization of Inference》研究任务信息与推理预算如何影响生产结果；《How Do Agent Harnesses Create Value?》考察规划和验证如何改变成功率、错误放行与成本；《Welfare-Opaque Income》进一步讨论，当 agent 的执行规则无法被政府观察时，税收政策如何判断真实的福利效果。

四篇论文采用不同的理论与实验方法，共同指向一个研究议程：理解 AI agent 如何改变经济活动中的行动权、价值创造和结果衡量，并据此设计更有效、更可问责的制度。

02 The papers系列论文

Four questions, one research agenda. 四个问题，一条研究主线。

The papers are ordered by their conceptual relationship. They use different theoretical or experimental methods, each with its own scope. 以下按照概念逻辑排列。各篇采用不同的理论或实验方法，结论有各自的适用范围。

01 Delegation rights委托权

Delegation Rights: Property, Agency, and Investment Incentives in the Age of AI Agents

Yukun Zhang · Kemu Xu · arXiv:2606.31935 [econ.EM]

Abstract · adapted摘要 · 据原文整理

AI agents can act through privileges already held by account users, raising a question about who controls the mode of execution. The paper defines delegation rights as revocable, scope-limited, identity-preserving, and mode-specific authorization. In a three-party incomplete-contracts model of the user, agent provider, and platform, platform discretion can discourage relationship-specific investment, while unrestricted user control can leave security, privacy, and third-party harms insufficiently addressed. The proposed certified-delegation regime protects access subject to verifiable authorization, auditability, and risk controls. Illustrative mechanism simulations suggest that this conditional allocation of control can improve delegation incentives while limiting residual risk. AI agent 可以借助用户已有的账户权限执行任务，由此产生一个新的控制权问题：用户能否授权自动代理以特定方式行使既有权利？论文将委托权界定为可撤销、限定范围、保留用户身份且区分执行方式的授权，并建立用户、agent 提供者与平台的三方不完全契约模型。平台自由否决可能削弱双方的投资激励，用户自由委托则可能留下安全、隐私和第三方风险。论文提出以可核验授权、审计和风险控制为条件的认证委托；示意性机制模拟表明，这种安排有可能在恢复委托激励的同时约束剩余风险。

Details, figures and BibTeX详情、配图与 BibTeX arXiv ↗ Hugging Face ↗
02 Production生产组织

The Organization of Inference: Information, Resource Constraints, and AI Production

Yukun Zhang · Kemu Xu · Yishen Chen · arXiv:2609.20449 [cs.AI]

Abstract · adapted摘要 · 据原文整理

The productive value of inference depends on where task information and computational capacity are placed within a workflow. In externally verified software-engineering tasks, direct execution succeeds in 59.6% of assignments at both 12,000- and 24,000-token ceilings, while information-constrained planning rises from 36.2% to 51.2%. A separate strict read-only experiment changes whether the planner sees the task issue: access adds about 16 percentage points at the lower ceiling, and task-informed planning exceeds direct execution by 29.6 points at the higher ceiling. These comparisons show that more capacity has different effects across workflows, depending on which stage receives information and how much resource remains for execution. 推理能力的经济价值取决于任务信息和计算预算如何在规划与执行之间分配。论文在经过外部验证的软件工程任务上比较不同工作流：逻辑 token 上限从 12,000 增至 24,000 时，直接执行的成功率均为 59.6%，信息受限规划的成功率则从 36.2% 升至 51.2%。在另一个严格的只读规划实验中，让规划者看到具体问题，在 12,000 token 条件下将成功率提高约 16 个百分点；在 24,000 token 条件下，任务知情规划比直接执行高 29.6 个百分点。结果说明，扩大预算是否有用，取决于信息能否到达需要它的阶段，以及执行阶段是否仍有足够资源。

Details, figures and BibTeX详情、配图与 BibTeX arXiv ↗ Hugging Face ↗
03 Execution control执行控制

How Do Agent Harnesses Create Value? Planning Information and Release Control in Stateful LLM Agents

Yukun Zhang · Kemu Xu · Yishen Chen · arXiv:2609.20474 [cs.AI]

Abstract · adapted摘要 · 据原文整理

Agent harnesses provide planning guidance and decide whether attempted completions should be accepted. The paper studies their effect on success, false passes, and cost in two Retail experiments and an Airline pilot from τ²-bench. Across 265 matched cells, prewritten task-specific guidance raises oracle-verified success by 7.17 percentage points relative to a word-count-matched shuffled control, with larger gains on more complex tasks. A read-only terminal verifier rejects 61% of invalid Retail episodes but also withholds 17% of correct ones. The preferred component depends on the loss assigned to false acceptance; when that loss is high, standalone verification can capture much of the risk benefit at lower cost. agent harness 通过提供规划指引和检查完成状态影响任务表现。论文使用 τ²-bench 的两组 Retail 实验与一组 Airline 先导实验，分别考察规划内容、终端验证、成功率及成本。在 265 个匹配单元中，预写的任务专属规划比等字数的乱序文本对照使 oracle 验证成功率提高 7.17 个百分点，收益主要出现在更复杂的任务。只读终端验证器拒绝了 61% 的 Retail 无效任务，也误拒了 17% 的正确任务。规划与验证的经济价值取决于错误放行的损失：损失较高时，单独部署验证器可能以更低成本取得大部分风险控制收益。

Details, figures and BibTeX详情、配图与 BibTeX arXiv ↗ Hugging Face ↗
04 Welfare福利衡量

Welfare-Opaque Income: Taxation under AI-Agent Delegation

Yukun Zhang · Kemu Xu · Yishen Chen · arXiv:2609.20425 [cs.CY]

Abstract · adapted摘要 · 据原文整理

When an AI agent implements economic choices through a rule the government cannot observe, the mapping from a person's preferences to execution becomes hidden alongside productive ability. The paper shows that identical observed tax-base responses can correspond to different welfare effects and derives an optimal-tax condition with an execution wedge: over- and under-execution change the local value of a marginal tax reform in different directions. In 4,500 controlled runs across five AI engines, faithful delegation almost always selects the score-maximizing option, while conflicting objectives produce different patterns of departure across engines. The laboratory motivates collecting execution information alongside tax-base statistics; it does not estimate real-world tax effects. 当 AI agent 按政府无法观察的规则替人作出经济选择时，除生产能力之外，偏好如何变成执行结果也成为隐藏变量。论文指出，即使可观察的税基反应和机械福利权重相同，税收改革的福利后果仍可能不同，并推导将执行偏差纳入其中的最优税条件：过度执行与不足执行会使提高边际税率产生方向不同的额外效应。论文还在五种 AI 引擎上进行 4,500 次受控模型运行。忠实委托几乎总能选中评分最优项，但目标冲突时，不同引擎偏离的方向与位置并不一致。结果提示，税基统计之外还需要理解代理执行规则；实验不直接估计现实经济中的税收效应。

Details, figures and BibTeX详情、配图与 BibTeX arXiv ↗ Hugging Face ↗
03 News动态
04 Authors作者

Yukun Zhang

The four papers in this series are coauthored by Yukun Zhang and Kemu Xu; three are also coauthored by Yishen Chen. For author order, original abstracts, and current versions, consult the individual arXiv records. 本专题的四篇论文由 Yukun Zhang 与 Kemu Xu 合作完成，其中三篇还与 Yishen Chen 合作。各篇论文的作者顺序、摘要原文及最新版本以对应的 arXiv 页面为准。

Yukun Zhang
The Chinese University of Hong Kong香港中文大学
215010026@link.cuhk.edu.cn
Kemu Xu
University of Edinburgh爱丁堡大学
s2749200@ed.ac.uk
Yishen Chen
The Chinese University of Hong Kong, Shenzhen香港中文大学（深圳）
yishenchen@link.cuhk.edu.cn
05 Research agenda研究议程

Making delegated action an economic object of study. 让委托执行成为可以研究的经济关系。

The four papers examine delegated action through rights, production, verification, and welfare measurement. Their models and experiments illuminate different parts of this research agenda. A common causal link from rights and internal workflows to external welfare outcomes remains a question for future research. 四篇论文分别从权利配置、组织生产、结果验证与福利衡量切入，构成相互关联的研究议程。各自的模型与实验揭示了不同环节；权利安排、内部工作流和外部福利结果之间的共同因果链，仍有待进一步检验。