Podcast Atlas
Source note Episode guide Original audio Topics: Technology

外滩大会线下圆桌｜敢把钱包交给AI吗？聊聊Agent交易爆发前夜的信任基建

Summary

This 硅谷101 roundtable at the 外滩大会, hosted by 红君, brings together 韩新毅, Yong Lambert, 刘作虎, and 周靖人 to examine what must exist before consumers will let agents transact. The panel’s shared thesis is that model capability and payment protocols are insufficient: Agent Payment Infrastructure / 智能体支付基础设施 needs authenticated users and agents, bounded authorization, verifiable intent, secure execution, traceability, recourse, and gradual trust calibration. Early usage is still thin, but merchant demand and reported shopping growth make the trust stack a present design problem rather than a distant hypothetical.

Key Claims

  • Only one panelist reports having completed a purchase through an agent, while OPPO’s source-scoped assistant-shopping data reportedly doubled from a small base, indicating early adoption rather than mature scale.
  • Agent payment must authenticate the user, establish that the executing agent is trustworthy, preserve the user’s actual instruction, and retain evidence for assigning responsibility when the outcome is wrong.
  • Verifiable Intent / 可验证意图 records the mandate behind a transaction so disputes can distinguish user ambiguity, agent misunderstanding, transmission failure, merchant error, and unauthorized action.
  • Agent Permission Boundaries and the harness layer must preserve strong constraints, use safe execution environments, and prevent later content from overriding budget or action limits.
  • KYC and KYB are no longer enough when machines act: the panel proposes KYA, or “know your agent,” alongside agent capability assessment and A2A behavior norms.
  • Consumer trust has both capability and risk dimensions. Users need evidence that an agent can perform the task and confidence that autonomy remains bounded, traceable, and remediable when something fails.
  • Agent-mediated recommendations still optimize among candidates using location, preferences, timing, and other objectives; hiding intermediate choices increases the importance of explainability, ranking integrity, and recourse.
  • Agent-to-agent commerce may require permissioned machine credentials, organizational authority, working capital, and machine-speed settlement rather than simply attaching today’s consumer card or wallet account.
  • The panel expects commerce distribution to move from attention and time toward intent, creating opportunities for personal assistants, intent-based merchant conversion, and scalable expert services.

Key Quotes

“可验证意图” — Yong Lambert’s label for preserving the user’s original transaction mandate.

“你敢付我敢赔” — Han Xinyi’s shorthand for recourse as a condition of delegated spending.

“流量会从时长变成意图” — Han Xinyi’s description of the proposed commerce-distribution shift.

Connections

Contradictions

  • No settled contradiction found. The forecast of a six-to-twelve-month agent-economy acceleration is explicitly balanced by the panel’s admission that prior agentic-commerce deployment was slower than expected and current user adoption remains small.
  • Reported shopping growth, adoption, forthcoming standards, blockchain suitability, and timing forecasts remain source-scoped; the translated and dubbed Yong Lambert remarks may not preserve every nuance of the original English.