AI 中文总结
研究平台承载第三方AI智能体时，开发者投资形成平台特定能力存量并面临套牢风险。通过建模源平台无法承诺条款、目的平台可验证识别规则的情形，刻画封闭、部分和完全可移植性，比较可移植性与费用承诺，指出验证的双重效应及最优可移植性条件，将智能体互操作性转化为关系能力的条件性治理选择。
AI 中文摘要
平台日益承载、分发并编排第三方AI智能体。开发者对工作流、工具连接、测试、任务记录和授权记忆的投资，会形成一种私有能力存量，并可能使互补智能体更具生产力。由此产生的存量在平台关系内部生成，具有平台特定性，且在投资后易被侵占。我们建模了一个无法承诺未来完全费用与路由条款的源平台、一个承诺可验证识别规则的目的平台，以及开发者先在源平台重置条款之前进行投资的情形。可移植性通过退出选项保护投资，同时保留源平台的运营裁量权。我们刻画了封闭、部分和完全可移植性，并表明投资结果在事后不完全抽取下依然成立。验证具有两种相反效应：它减少了转移损害，但可能强化目的平台的识别能力和开发者的外部租金。仅当安全性和诱导投资收益超过额外租金与竞争泄漏时，验证才会扩大最优可移植性。在给定目标保护回报的条件下，我们将可移植性与费用承诺进行比较，并刻画了成本最低的工具组合。最后，当私人泄漏主要是再分配性质时，源平台过于封闭；但当重复、碎片化或安全成本足够真实时，源平台可能过于开放。该分析将智能体互操作性转变为一种关于关系能力的条件性治理选择，而非预设开放性总是有效。
英文摘要
Platforms increasingly host, distribute, and orchestrate third-party AI agents. A developer's investment in workflows, tool connections, testing, task records, and authorized memory creates a private capability stock and can also make complementary agents productive. The resulting stock is generated within a platform relationship, becomes platform specific, and is vulnerable to appropriation after investment. We model an origin platform that cannot commit to complete future fee and routing terms, a destination platform that commits to a verifiable recognition rule, and developers that invest before the origin resets terms. Portability protects investment through an exit option while preserving the origin's operating discretion. We characterize closed, partial, and full portability and show that the investment result survives incomplete ex post extraction. Verification has two opposing effects: it reduces transfer harm but can strengthen destination recognition and developer outside rent. It expands optimal portability only when safety and induced-investment gains exceed the additional rent and competitive leakage. Conditional on a target protected return, we compare portability with fee commitment and characterize the least-cost instrument mix. Finally, the origin is too closed when private leakage is mainly redistributive, but can be too open when duplication, fragmentation, or security costs are sufficiently real. The analysis turns agent interoperability into a conditional governance choice over relational capability rather than a presumption that openness is always efficient.
Comments65 pages, 4 figures, including a 30-page Online Appendix with proofs, robustness analyses, a bounded software-policy audit, and a reproducibility map
