从 Harvey 毛利到私人助手：代理真正要完成的，是一份工作而非一个动作

A temporary collapse in gross margin can reveal a mismatch between pricing and agent usage, but the subsequent response matters as much as the alarming snapshot.

Sustainable agent products need task-level cost visibility and spending controls, alongside model routing and post-training that preserve the quality customers are paying for.

Human support can help an early agent service finish difficult tasks, while raising separate questions about scalability, disclosure, privacy, and whether any resulting data may be used for training.

A small private workflow tool can create substantial value by capturing requests reliably and delivering them at the right time, without replacing the person who performs the work.

Automating one task does not complete the whole job when the remaining work still includes physical handling, exception management, and confirmation that the user's actual goal was achieved.

Consumer agents may be easier to adopt when they relieve specific annoyances and preserve enjoyable choices, rather than treating every moment of research or shopping as wasted productivity.

An agent can reorganize product discovery without automatically reproducing the logistics, returns, and delivery reliability that make customers choose an established commerce platform.

An educational program for builders can prepare students for several meaningful career paths, rather than measuring every student's success by whether they become a founder.

As models improve, valuable training data increasingly targets specific weaknesses and difficult contexts, requiring a blend of human expertise and specialized AI rather than indiscriminate volume.

A physical product's viability depends on its manufacturing process, operating requirements, and cost structure as well as its performance, so these constraints cannot simply be added after the prototype works.