将对话转化为知识：Slack 如何构建人机 Agent 团队
英文标题：Turning conversation into knowledge: how Slack builds human-agent teams
Slack 首席产品官 Jaime Delanghe 认为，要让工作场景中的对话沉淀为组织知识，团队需要默认在公开频道中协作，让人类与 Agent 共享可搜索的上下文。她建议通过由 Agent 生成的简报、清晰的角色分工、共享频道中的任务交接，以及可复制的实践分享，建立有效的人机 Agent 协作循环。Slack 的经验表明，衡量 AI 价值不能只看 Token 使用量或活跃度，而应关注实际工作成果。对开发者和企业而言，关键不只是用 AI 加速既有流程，更是以团队方式重新设计工作流，并从小规模试点开始推广。
19.0 KB
403 B
10.5 KB
10.2 KB
11.6 KB
13.4 KB
62.7 KB
中文
这是我们“构建人机 Agent 团队”系列的第二篇文章。第一篇分享了我们在 Anthropic 构建多人协作 AI 团队时学到的经验。本文将分享一家早在 AI 出现之前，就已开始思考人机 Agent 团队的公司的最佳实践。
Jaime Delanghe 于 2017 年加入 Slack，负责搜索和机器学习，使命是将工作场所中的对话转化为机构知识。如今作为公司的首席产品官，她从一开始就相信：要实现这一目标，人们需要以开放的方式协作，将对话、决策和进行中的工作保留在公司内任何人都能阅读和搜索的频道中。在她最近发表的文章《工作即对话》中，她也对 Agent 提出了同样的观点：围绕工作的对话，是 Agent 发挥作用所需的上下文，并最终能帮助我们实现这项延续数十年的目标——把分散的知识转化为生产力。
为了了解这一模式在 Slack 中的实际运作方式，我们与 Jaime 进行了交流，讨论她在构建有效人机 Agent 团队以及推广这些新工作方式方面的最佳实践。
将你的对话历史视为知识库
多年来，人们一直期待工作场所中的对话——即人们协作时产生的“尾气”——能够累积为组织知识，但这一承诺始终没有实现。
Jaime 表示：“我收藏了许多 Slack 早期的研究论文，它们实际上表明：不，对话并不会转化为知识。你希望它会，但实际上，那里只是堆着大量内容，人们仍然不得不反复解释同样的事情。”
人类根本无法理解和梳理所有这些“尾气”。如今，这是 Agent 的工作。
如何将其付诸实践
- 默认使用公开频道： Agent 只能从它们看得到的内容中学习。在私信或私密讨论串中做出的决定对 Agent 不可见，也会继续遗失在组织之外。
- 要求 Agent 提供推理过程，而不只是记录： 与其搜索最终做了什么决定，不如要求 Agent 重建_为何_做出这一决定，以及此后上下文发生了怎样的变化。
- 扩大上下文覆盖面： Slack 和 Claude 等工具正在将会议、电子邮件、日历和文档库连接起来——你接入的上下文越多，团队重复劳动就越少。
学会何时在 Agent 与人类之间交接任务
人机 Agent 团队的核心节奏，是一个持续交接的循环。由 Slack 中的 Claude 驱动，Agent 负责生产性工作——起草、总结、监控、准备——再将结果交给人类。人类负责审阅、决策和重新引导，然后将工作交回 Agent 执行下一步。
若想看到这一模式的实际应用，看看 Jaime 如何开始自己的一周即可。
Jaime 说：“现在是周一早晨，我刚刚查看完一份 Agent 为我制作的每日简报。”
等待她审阅的还有：上周产品研讨会的回顾及标记出的升级事项、全网 AI 发展报告、当天会议的简报，以及她此前交给 Agent 重写的一份过时个人简介。每个循环结束时，人类都会审阅，并根据 Agent 的行动做出决策。
如何将其付诸实践：
- 以 Agent 制作的简报开启一天。 回顾、升级事项、会议准备和网络信息汇总，都是适合由 Agent 推动、再由人类审阅的任务。
- 将工作锚定在共享频道中。 在共享频道中分享全部工作，让人类和 Agent 能共同分流处理，其中人类主导优先级排序。
- 让轻量信号可以直接触发行动。 在 Jaime 的频道中，一个表情回应就能将某项内容加入列表，随后由 Agent 接手任务。
为 Agent 分配清晰角色
如果你的思维模型仍是与单个聊天机器人一对一互动，那么与一组专门化 Claude Agent 协作可能会令人无所适从。Jaime 的方法更偏社会协作，而非技术视角：“我喜欢把 Agent 看作有点像同事。”
就像人类队友拥有角色和职责一样，Agent 也应有清晰的目标和关注领域。她说：“如果 Agent 的价值更像是被强制要求，而不是被使用者清晰地感受到和理解，那么人们就很难记住它到底是做什么的。”
如何将其付诸实践：
- 将常规、事务性任务交给通用 Agent。 不要要求人们记住某个专用工具，而应训练一个 Agent 处理重复任务，例如提交帮助台工单，或将上周指标提取到状态更新中。
- 让价值被感受到，而不是被强制规定。 如果人们无法说明一个 Agent 的用途，也许该考虑将其淘汰。
默认将共享频道设为公开；有目的地设为私密
Slack 从早期开始就建议频道默认公开。Jaime 说：“你是在为之后所有工作构建一种共享理解和共享上下文。”
她建议，除非存在需要限制上下文和知识访问的具体原因，否则应保持频道公开。Agent 能够提取并用于工作的资料越多，它们作为团队伙伴就越有效。
开放的上下文会不断累积——新成员接触的是历史记录，而不是空空如也的收件箱；也没有人需要反复解释同样的事情。现在 Agent 也能从中受益，而这些上下文和工作记忆又会回流给人类。
如何将其付诸实践：
- 将日常业务工作置于开放环境。 将非敏感项目、公告和问答频道设为公开，让 Agent 同事获取最有助于发挥作用的知识。
- 记住，你的 Agent 能阅读团队成员所阅读的内容。 对每一个需要从私密频道获取信息并进行汇报的 Agent 而言，私密频道都是盲区。
- 让心理安全感决定边界。 一旦真正敏感的材料被隔离，工作转入私信的主要原因通常不再是保密，而是不愿在过程尚未完成时被他人看到。人们应有信心在公开环境中进行日常工作，包括粗略草稿和尚未成形的问题，并相信同事会以善意回应。这种开放会不断累积：“给予信任，才能获得信任。”
通过展示可能性来推动采用
学习一种新工作方式最快的方法，是观察队友如何操作。Jaime 在 Salesforce 见过这种情况：员工会在一个名为_How I Slackbot_的全公司频道中分享技能、调试技巧和工作流窍门。按她的估计，该频道拥有数千名成员。这个默认公开的频道中，销售流程中的一个技巧，可能最终重塑工程流程。
在 Slack 内部，推动产品经理使用 Claude 的行动“是你能想象到的最自组织化的事情”。一位产品经理请开发者体验负责人帮助自己完成设置，随后他写了一份 canvas，展示自己做了什么以及如何做到。其他产品经理复制了这种格式。各团队组织研讨会，并建立了自己的 git 仓库。
如何将其付诸实践：
- 建立一个全公司范围的展示交流频道。 为员工提供一个公开场所，用于分享技能、调试技巧和工作流窍门，让一个职能中的技巧能够重塑另一个职能。
- 鼓励撰写可供他人复制的说明文档。 一份简短的“我做了什么、如何做到”文档，可以将一个人的配置方式转化为团队模板或技能。
衡量结果，而不是活动量
从加入 Slack 的早期开始，Jaime 就一直在思考如何衡量生产力。她说：“我们是希望人们发送更多消息吗？也许不是。发送更多消息并不一定意味着他们从 Slack 中获得了更多价值。更多消息可能意味着人们找不到所需内容，或者无法第一次就准确表达自己的意思。”
如今，衡量 AI 价值的问题看起来非常相似——面对如此复杂的事物，简单指标无法胜任。Token 使用量可以告诉你系统是否正在运行，但虽然这一点很重要，却远远不够。
如何将其付诸实践：
- 将使用指标视为健康检查，而非价值证明。 活动量只能说明采用正在发生，不能说明它确实有效。
- 准备好运用自己的判断。 没有一种清晰的方法能够证明，人们使用这些工具的方式一定会带来更好的业务结果。正如 Jaime 所说，将两者联系起来仍需要“许多信念上的跳跃”，没有任何仪表盘或使用统计数据能替你证明这一点。
共同改变工作方式
对于试图实施人机 Agent 团队的组织，Jaime 最重要的建议是重新构想每一项工作流：“我们必须想办法改变我们的工作方式，而不只是更快地完成更多同类工作。这将是一项团队运动。”
她对构建有效人机 Agent 团队的最大建议是什么？尽早开始，但从小处着手。将一群人带入一个与 Claude 共享的频道，为他们提供同一套资源，然后让他们开始工作。如果 Slack 的经验具有参考意义，那么他们构建出的实践会自行传播。
English / 原文
This is the second post in our series on building human-agent teams. The first shared what we've learned building teams with multiplayer AI at Anthropic. In this article, we share best practices from a company that was thinking about human-agent teams long before AI arrived.
Jaime Delanghe joined Slack in 2017 to work on search and machine learning, with a mission to turn workplace conversation into institutional knowledge. Now the company’s Chief Product Officer, she has believed from the start that to achieve this goal, people need to work in the open, keeping conversations, decisions, and work in progress in channels anyone at the company can read and search. In her recent essay The Work is the Conversation, she makes the same case for agents: The conversation around the work is the context that agents need to be useful and finally help us achieve this decades-old goal of turning scattered knowledge into productivity.
To learn what this looks like in practice at Slack, we talked with Jaime about her best practices for building effective human-agent teams and spreading these new ways of working.
Treat your conversation history like a knowledge base
For years, the promise that workplace conversation—the "exhaust" of people working together—would compound into organizational knowledge never materialized.
"I have so many research papers from the early days at Slack that showed that, actually, no, conversation doesn't turn into knowledge," Jaime says. "You wish it did, but really it's just a lot of stuff that just hangs out there and people still have to repeat themselves."
Making sense of all that exhaust simply wasn't humanly possible. Now it's an agent's job.
How to put this into practice
- Default to public channels :**** Agents can only learn from what they can see. Decisions made in DMs or private threads are invisible to them—and stay lost to the organization.
- Ask agents for the reasoning, not just the record : Instead of searching for what was decided, ask an agent to reconstruct why it was decided, and how the context has shifted since.
- Widen the surface area : Tools like Slack and Claude are stitching meetings, emails, calendars, and document repositories together—the more of that context you connect, the less your team repeats itself.
Learn when to handoff tasks between agents and humans
The core rhythm of a human-agent team is a cycle of handoffs. Powered by Claude in Slack, agents handle the production work—drafting, summarizing, monitoring, preparing—and pass the results to a person. The person reviews, decides, and redirects, then hands the work back for agents to carry out the next step.
To see all this in practice, look no further than how Jaime starts her week.
"It’s Monday morning, and I’ve just had my daily briefing that an agent has built for me,” Jaime says.
Also waiting for her review is a recap of the previous week's product workshops with flagged escalations, a report on AI developments across the web, briefings for the day's meetings, and a stale bio she'd handed to an agent to rewrite. At the end of each loop, humans review and make decisions based on the agent’s actions.
How to put this into practice:
- Start the day with agent-built briefings. Recaps, escalations, meeting prep, and web roundups are great tasks for agents to drive, with human review.
- Anchor the work in a shared channel. Share all work in a shared channel so that humans and agents can triage it together, with humans leading the charge on prioritization.
- Make lightweight signals actionable. In Jaime's channel, an emoji reaction adds an item to the list and an agent picks up the task.
Delegate clear roles for agents
Working with a fleet of specialized Claude agents can feel disorienting if your mental model is a one-on-one chatbot. Jaime's approach is social rather than technical: "I like to think that agents are kind of like coworkers."
In the same way that human teammates have roles and responsibilities, agents should also have clear goals and focus areas. "If the value of the agent feels mandated rather than very clearly felt and understood by the people using it, it's really hard to remember what the thing is for,” she says.
How to put this into practice:
- Route routine, transactional tasks to a general agent. Rather than asking people to remember a specialized tool, train an agent to tackle a repetitive task, like filing a help desk ticket or pulling last week's metrics into a status update.
- Let value be felt, not mandated. If people can't articulate what an agent is for, it may be time to retire it.
Default shared channels to public; go private on purpose
Slack has recommended public-by-default channels since its earliest days: "You're building a shared understanding, a shared context for all of the work that's going to come next,” Jaime says.
She suggests keeping channels public unless there is a specific reason to gate context and knowledge. The most information agents have to pull from and inform their work, the more effective team mates they’ll be.
Open context compounds—new people onboard into history instead of an empty inbox, and no one repeats themselves. Now agents benefit too, and that context and working memory flows back to humans.
How to put this into practice:
- Keep business-as-usual work in the open. Make**** non-sensitive**** projects, announcements, and Q&A channels public so that agent coworkers can gain the knowledge they need to be most useful. .
- Remember your agents read what your team reads. A private channel is a blind spot for every agent that reports on it.
- Let psychological safety drive the line. Once genuinely sensitive material is walled off, the main reason work retreats into DMs isn't secrecy—it's discomfort with being seen mid-process. People should feel confident doing everyday work in the open, rough drafts and half-formed questions included, trusting their coworkers to meet it in good faith. And that openness compounds: "you gain trust by giving trust."
Spread adoption by showing the art of the possible
The fastest way to learn a new way of working is to watch a teammate do it. Jaime has seen this at Salesforce, where employees share skills, debugging tips, and workflow tricks in a company-wide channel called How I Slackbot , which by her count has thousands of members. In that channel, which is public by default, a trick from a sales process can end up reshaping an engineering process.
Inside Slack, a push to get product managers using Claude "was the most self-organized thing you could possibly imagine." One PM got the developer experience lead to help him get set up, then he wrote up a canvas showing what he did and how he did it. Other PMs copied the format. Teams organized workshops and built their own git repos.
How to put this into practice:
- Stand up a company-wide show-and-tell channel. Give employees one public place to share skills, debugging tips, and workflow tricks, so a trick from one function can reshape another.
- Encourage write-ups others can copy. A short "what I did and how" doc turns one person's setup into a team template or skill.
Measure outcomes, not activity
Since her early days at Slack, Jaime has grappled with the question of how to measure productivity. "Do we want people to send more messages?” she says. “Maybe not. Sending messages might not actually mean that they're getting more out of Slack. More messages can mean people can't find what they need, or can't say what they mean the first time."
Now, the question of measuring the value of AI looks quite similar—and with something that complex, simple metrics don’t do the job. Token usage tells you the lights are on, but while that’s important to know, it’s not sufficient.
How to put this into practice:
- Treat usage metrics as a pulse check, not proof of value. Activity tells you adoption is happening, not that it's working.
- Be ready to use your own judgment. There's no clean way to prove that how people use these tools leads to better business results. As Jaime puts it, connecting the two still takes "a lot of leaps of faith," and no dashboard or usage stat will prove it for you.
Change how you work, together
Jaime's biggest piece of advice for organizations trying to implement human-agent teams is to reimagine every workflow: "We're going to have to figure out how to change the ways that we're working, not just do more of the same kind of work faster. And that is going to be a team sport."
Her biggest advice for building an effective human-agent team? Start soon, but start small. Bring a group of people into a shared channel with Claude, give them the same set of resources, and let them work. If Slack's experience is any guide, what they build will spread on its own.