E255｜模型越来越强，为什么用户没感觉？再访阿里国际站总裁张阔

Summary

This 硅谷101 follow-up interviews 张阔 / Zhang Kuo about why stronger general models do not automatically produce equally visible gains in real commerce. Using Accio Work and a 107-task ecommerce evaluation, the episode argues that a production vertical agent depends on model, harness, and business context together, while open-ended goals, cross-system execution, long feedback cycles, and human accountability create a persistent Business Agent Benchmark Gap. The practical boundary is a governed division of labor: agents can research, organize, monitor, publish, and execute repeatable work, while people retain product taste, resource allocation, payment approval, and the overall business objective.

Key Claims

• Coding agents benefit from digital inputs, abundant training data, compilers, tests, and fast feedback; commercial agents face heterogeneous systems, organization-specific goals, open-ended outcomes, and feedback that may arrive months later.
• Accio Work is presented as an expansion from multimodal procurement search into market research, product design, supplier matching, communication, logistics, multi-platform publishing, pricing, customer feedback, and operating analysis.
• 张阔 / Zhang Kuo describes agent capability as “model × Harness × Context”: tools, memory, permissions, runtime continuity, and proprietary business data are complements to model intelligence rather than optional wrappers.
• The team reportedly derived 107 ecommerce tasks from real operating and product-question data; the source says frontier models complete about 61% without human intervention even while some general benchmarks approach 99.
• The benchmark spans supplier-quote screening, landed-cost calculation, shipping booking, dispute handling, and other tasks across four difficulty levels, making successful execution—not a plausible answer—the pass condition.
• Model Routing Cost Control sends routine work to smaller or cheaper post-trained models and reserves stronger frontier models for harder tasks or modalities so the product remains affordable at operating scale.
• Cloud continuity alone is not Long-Horizon AI: a useful commercial agent must retain goals, incorporate new supplier and sales information, recover across interruptions, and produce verifiable business outcomes over weeks or quarters.
• Human approval remains mandatory for consequential actions such as purchase payment, while product need, customer willingness to pay, inventory and advertising allocation, taste, and market timing remain management judgments.

Key Quotes

“模型 × Harness × Context” — Zhang Kuo’s compact description of the three complementary layers behind a vertical agent.

“商业领域的AGI” — defined in the episode as an agent producing more operating return than a person under the same budget and supply-chain conditions.

Connections

• 张阔 / Zhang Kuo, Alibaba, Accio Work, and 硅谷101 — guest, company, product, and show behind the follow-up.
• Vertical Agent Architecture, Agent Harness, Context Engineering, and Enterprise Agent Production Environment — model-external architecture needed for real commercial execution.
• Business Agent Benchmark Gap, Agent Evaluation Benchmarks, and Environment-Based Agent Benchmarks — difference between general scores and verified work completion.
• Agentic B2B Sourcing, B2B to A2A, AI As Business Operator, and Ecommerce Fulfillment Complexity — procurement and operating workflows covered by the product.
• Model Routing Cost Control, AI Inference Cost Structure, and Agent Post-Training — capability-cost allocation across tasks.
• Long-Horizon AI, Persistent Cloud Agents, Persistent Agent Memory, and Agent Permission Boundaries — continuity, state, authorization, and human control.
• Human Judgment Under AI and Human Agency Under AI — retained roles in goal choice, taste, resource allocation, and accountable decisions.

Contradictions

• No settled contradiction is recorded. The source strengthens E231’s claim that cross-border business agents require workflow-specific data, verification, permissions, and long-lived state rather than a general chatbot.
• The 20% search growth, 130% multimodal-search growth, 24-hour sourcing example, 61% benchmark pass rate, event attendance, and entrepreneur revenue figures are product-provider claims without independent validation in the supplied episode.
• The episode qualifies benchmark pessimism: 61% refers to complete tasks without human intervention and does not establish that assisted workflows lack value.
• It also qualifies cloud-agent narratives: staying online or synchronizing memory is infrastructure, not proof that an agent can preserve a business objective or deliver a correct long-term result.