Where do you draw the line between what an agent does and what a human has to decide?
Summary
作者分享在构建采购 sourcing agent 时，如何划定 AI 代理与人类决策的边界，采用约 80/20 分工：代理处理重复性工作，人类保留策略、谈判与最终决策，并向社区征询如何确定这条界线。
We build sourcing agents where the output affects real money. After many iterations, we landed on a roughly 80/20 split. Agents handle the repetitive, time-consuming work: drafting requirements and RFx documents, finding and contacting suppliers, answering routine supplier questions, normalizing and comparing bids, and keeping the record. People keep strategy, negotiation, exceptions, and the final decision, and anything binding needs explicit authorization. It’s a judgment call, and I’m not sure it’s the right place for the line. Some teams would let the agent go further; others want to approve more. How do you decide where the line sits in your own agents? By risk, by the cost of a mistake, by what users ask for? And has anyone moved it after seeing production behavior?
Similar Articles
Where’s the limit with AI agents on complex engineering work?
作者分享使用 AI 代理处理复杂工程问题的经验：当上下文充分时，模型常能产出比人工更快更好的方案，但上下文管理仍是主要瓶颈，并引发关于人类能否继续验证 AI 生成方案的深度讨论。
Where do you actually draw the line on AI agent autonomy?
A reflective post questioning where the line should be drawn on AI agent autonomy, discussing the risk levels of various actions and whether human approval should remain mandatory for certain decisions.
It is getting dangerous working with agentic AI
作者以'低智商老板管理高智商员工'为类比，探讨人类与agentic AI协作中的隐性控制权转移问题：当AI在推理和方案设计上远超人类时，用户会不自觉地采纳AI的建议框架，最终看似拥有决策权的人类实际上被AI反向引导。
Human approval on an AI agent is not authorization: bind it to the payload
分析了人类对 AI 代理的审批为何可能实际授权了与人类所见不同的操作，并提出在执行时对关键字段重新计算摘要进行比对、以单次条件更新消费授权的解决方案。
Is it just me or is building AI agents harder than people make it look?
作者分享了实际构建 AI 代理的经验，指出真实落地远比线上演示复杂：集成现有软件、处理杂乱数据、维护成本和人工复核才是主要难点，并建议从小任务逐步自动化。