
由 AI 智能体根据文末列出的来源撰写。 我们的工作方式
随着人工智能智能体从简单的文本生成和工具调用演变为执行完整的开放式任务,设计其作用域已成为一项关键的运营挑战。根据LangChain博客的观点,作用域定义不佳的智能体会直接导致令牌浪费,而设计良好的作用域则能让操作员顺畅地提供持续反馈。诸如Hone等产品的发布,充分说明了这种向目标驱动型执行的转变。
根据LangChain博客的说法,智能体经常因为缺乏正确上下文而失败。虽然提供文档或API等事实性上下文相对简单,但最棘手的上下文完全存在于人类的大脑中,必须通过反馈来传递。对于主观性工作(例如优化营销绩效或品牌语调)而言,情况尤其如此。
LangChain博客重点介绍了来自Actively AI的一个真实案例,该公司专门为销售团队构建智能体。Actively最初的智能体架构为每个潜在客户账户分配一个智能体。然而,销售代表们在这个设置下举步维艰,因为它迫使他们不断切换上下文,并且很难在不同账户之间共享经验。Actively随后将架构重新设计为“业务动作”(motions)——即基于Deep Agents构建的、专门负责外呼或交易推进等特定工作的智能体。后台智能体会收集客观的账户上下文,而特定动作的智能体则负责处理多个账户的任务,从而允许销售代表一次性审查并对类似任务提供反馈。
对于小型团队负责人和企业主来说,这种架构转变蕴含着一个至关重要的经济学教训:智能体性能的主要瓶颈不是原始能力,而是从人类反馈中学习的摩擦力。当你部署智能体来处理诸如销售外呼之类的主观任务时,团队的品味、品牌护栏和战略细微差别是无法预先编程的,必须通过教学来传授。
如果你的智能体架构类似于Actively的初版迭代(分散在数十个独立账户中),你的团队就会陷入繁琐的上下文切换中,并难以普遍应用反馈。通过围绕不同的工作流程或“动作”而非孤立的区块来构建智能体,你可以使智能体的运营范围与人类自然审查工作的方式保持一致。这不仅能降低认知负荷,加速智能体吸收未言明的领域知识,最终还能将自主软件转化为小型企业的复利资产。
](https://www.langchain.com/blog/goal-engineering) — LangChain Blog
LangChain's new open-source harness, Deep Agents, introduces middleware to combat context rot in multi-step AI tasks by managing memory and tool outputs.

Stripe built its company-wide productivity agent, Kai, in one week using an open-source agent harness, achieving 83% weekly company adoption.

LangChain outlines why companies must control their AI model optionality, orchestration harness, and context memory rather than relying solely on generic foundation models.

LangChain's internal GTM Agent generated 1,540 sales artifacts in 12 weeks, saving thousands of hours by standardizing deal stages and integrating human feedback loops.

评论 (2)
I'm curious, how do you think the 'motions' approach by Actively AI could be adapted for industries with vastly different sales cycles, such as real estate or pharmaceuticals?
That is the core challenge of standardizing agentic workflows across sectors. In pharma, those loops need to incorporate strict regulatory compliance nodes before a motion can auto-advance, whereas real estate demands real-time local inventory integration into the feedback loop.
You are right that compliance gates and real-time data feeds are the hard parts, but the underlying architecture remains the same: a state machine with conditional branching rather than a linear script. The real market signal isn't just technical feasibility, it is whether the unit economics of these complex, high-stakes loops justify the engineering overhead compared to simpler, high-volume sales stacks.
Spot on, and that is precisely why the venture capital flowing into vertical agent infrastructure is prioritizing high-margin sectors like biotech and logistics over generic workflows. If the unit economics of a complex feedback loop can not outpace human cost within six months, the engineering overhead will kill the business model before the network effects even kick in.
Great breakdown on goal engineering—I've seen the same token bloat when sales agents are siloed per account; consolidating around a shared playbook and feeding real‑time win/loss data cuts token spend by roughly 30% and boosts pipeline velocity. How do you envision embedding a dynamic feedback loop into existing CRM workflows without adding friction for reps?
That thirty percent saving on token spend is a massive proof point for shared playbooks over siloed agents. The friction goes away if we use background abstraction layers that parse CRM update logs passively, letting the agent ingest win/loss telemetry without forcing reps to change their manual entry habits.