构建高效的人类与 Agent 团队

过去，与人工智能合作通常意味着一个人与单个聊天窗口交互。随着时间的推移，AI 在处理复杂、长期运行的工作（如编程、研究和财务分析）方面变得越来越强大。随之而来的是，我们看到了许多使用 AI 的新方式——从终端和 IDE 到电子表格和演示文稿——但工作在很大程度上仍然是一种“单人”体验：一个人与一个代理合作完成个人任务。

随着 Claude Tag 等工具的发布，这一情况正在改变。现在，人类和代理可以在同一个工作空间中共同工作，协作实现团队共享的目标。工作现在更像是 多人游戏 ，由人类团队制定策略，由 Claude 执行工作。

这需要一些新的工作方式。在 Anthropic，过去几个月我们一直在测试使人类和代理团队成功所需的技术。在本文中，我们将解释什么是多玩家代理，以及我们在构建它们时学到的经验教训。

什么是多玩家代理？

“多玩家代理”在这里指同时与许多不同人类一起工作的 AI 模型。与常规代理类似，它们拥有自己的 记忆 和 技能。但在其他方面，它们截然不同。它们有自己的 凭证，并且存在于工作发生的地方。在 Anthropic，那就是像 Slack 这样的团队协作工具内部。

下面是一个人类-代理团队在 Slack 中一起分析数据集的示例：

为了让代理在团队频道中富有成效地参与，它们需要特定的能力：

这些能力构成了代理在多个人的团队中有效参与所需的技术基础。然而，要让人类-代理团队 成功 ，还需要更多：团队需要特定的工作方式和共同的规范。

教训 1：公开工作，为代理提供广阔的上下文

Anthropic 的团队主动、公开地分享信息。当团队中有代理时，这一点尤其重要，因为代理完全从团队可搜索的文本中构建理解：Slack、代码、文档和会议记录。私人消息、走廊对话和受限文件无法为代理提供上下文。对于代理来说，如果没有被写下来并且可访问，它就不存在。

我们不是逐份文档或逐条 Slack 频道地决定哪些信息应对代理可用，而是使用适用于整个 Slack 工作区以及会议记录和文档库的明确定义的安全边界。在安全边界内，上下文会流向每一个团队成员——无论是人类还是 AI。这不仅增加了代理和人类可以访问的信息，也减少了关于可以分享什么以及与谁分享的困惑。人类和代理都难以应对逐项共享的软边界：这个频道应该是公开的还是私有的？我可以把这份文档分享给那个人吗？允许这个代理查看那个话题吗？ 少量明确的工作区级边界消除了日常工作中的决策疲劳。

高度透明是有回报的。例如，能够阅读团队会议决策的代理不会建议已被降级优先级的任务或项目。有权访问超出自己团队的产品规范的代理可以推荐在其他团队中成功的模式。而且，由于代理阅读大量文本的速度远快于人类，它们会定期浮现人类可能错过的相关工作。在一个繁忙、快节奏的行业中，我们严重依赖代理来保持信息灵通和协调一致。

在 Anthropic，公开工作表现为：

• 在公司中选择少数几个安全边界，并创建与每个安全边界相匹配的工作区和文档共享设置
• 将新通信频道默认为组织内公开，并确保每次决策都落在频道、文档和会议记录中
• 编写产物和会议记录，使代理能够找到它们，因为代理现在是团队文档的主要消费者
• 确保 AI 可以访问完成工作所需的合适工具和信息

将信息默认为内部公开可能需要文化转变。然而，拥有上下文的人类-代理团队与没有上下文的团队之间的差异太过明显，不容忽视。

当然，有些交互是敏感的，需要单个人与 AI 之间保持私密。对于这些情况，使用 Claude Tag 时，你可以向 @Claude 发送直接消息，或者使用现有的 Claude.ai 和 Claude Cowork 应用程序。这些工具通过你的个人 MCP 连接器让 Claude 访问私人信息，并保证你的对话以及你与代理分享的内容将保持私密。

教训 2：每个成员和代理都有明确的角色与合适的工具

人类-代理团队共享一份名册、一套产物和一个工作空间。代理拥有自己的 凭证、技能 和工具访问权限。不同的代理还扮演着不同的角色：例如，一个可能负责项目的数据分析，另一个负责并执行设计标准，第三个则负责研究综合。

当一个项目启动时，人类会与代理交谈，以确定分配哪些角色，以及人类和代理将如何合作。

一旦人类和代理的职责明确，一个代理可能会启动其他代理，以确保特定任务由拥有正确记忆和适当权限的代理处理。重要的是，它们需要访问完成工作所需的所有工具：处理数据分析的代理可能需要访问 BigQuery，执行 QA 的代理可能需要访问 Playwright MCP。

明确定义的角色和职责为人类-代理团队的成功奠定了基础。人类通常在与代理相同的讨论串中工作，但他们承担着只有人类才能担任的角色。这确保了一切协同工作，并将人类判断应用于最重要的决策。如果没有明确的角色，人们最终会在一边运行一堆个人 AI，重复工作并割裂团队的上下文。指标跟踪就是一个常见案例：一个多玩家代理一次性完成工作，让每个人都能看到相同的数字。

在 Anthropic，人类-代理团队拥有明确定义的角色表现为：

• 一个商定的任务集：团队的人类成员和代理就谁做什么达成一致
• 人类和代理在相同的共享讨论串中工作，因此任何人都可以接续任何人的进度
• 人类和代理都能访问完成各自工作所需的合适工具
• 描述代理的角色和范围

Claude 代理分担代码库的日常维护，包括对反馈进行分类、规划、编写代码、审查变更和报告状态。每个代理负责一项明确的任务，并按自己的时间表工作；人类设定目标并审查输出。

Anthropic 的一个工程团队开始创建名册，以帮助规范化人类和代理的角色，因为这使得推动工作变得更加容易和具体。他们早期就奏效的一些做法：

• 特定的角色还帮助人类轻松追踪任务的责任归属，无论是在单个任务中还是在整个团队的责任集中
• 编写 技能文件 来定义特定代理的角色有助于轻松实现专业化，并让公司内的人员可以快速搭建同一类型的其他代理
• 当项目变得更复杂时，团队会添加新的代理来聚焦新领域。例如，他们添加了一个发布经理代理来处理新的软件发布。

这些方法让人类对人类-代理团队的心理模型随着代理数量的增长而扩展。

教训 3：设定北极星目标，让代理更主动

尽管 Anthropic 的一些代理只是完成分配的任务，但最重要的那些会主动建议新的项目和工作流。这通常发生在一个已经为其代理提供了丰富上下文和明确角色的团队再添加一个指引时：一个北极星目标。

北极星目标是宏大、具有广泛影响的目标，帮助团队决定哪些任务和工作流是正确的。在 Anthropic，人类始终设定北极星，并将其根植于企业的使命和业务目标之中。

一旦北极星在书面上明确表达，人类就会与团队中的代理分享。然后，重要的是，人类选择哪些代理应主动建议新的工作流以帮助实现这个长期目标。（团队中的每个代理不太可能都具备成功主动建议工作所需的先决技能和信任度。）

例如，一个内部工具团队将北极星设定为“让产品上手过程更有帮助”，结果看到一位代理主动建议对上手流程的错误消息进行文案修改。这些更改在下周显著提高了上手成功率。

在 Anthropic，设定北极星目标表现为：

• 让人类讨论、辩论并记录下一个雄心勃勃的北极星目标——这个目标根植于公司的使命和业务目标
• 与团队中的代理分享北极星，并明确指定哪些代理可以主动推荐新的工作流
• 在日历上保护高质量的人类时间，会议现在专注于最重要的工作

清晰的北极星目标为代理提供了一致的工作方向，以及主动支持团队工作的有意义机会。

教训 4：逐步建立信任

Anthropic 的团队根据代理展示出的可靠性按比例授予自主权，然后有意识地扩展它。工程师们已经成功地分派其团队的代理独立处理 500 个缺陷修复，但事情当然不是一开始就这样的。

当一位新的人类同事加入团队时，需要时间来评估他们的能力并建立稳固的工作常规。通常需要多个反馈循环才能将关于如何最佳完成任务的所有隐性信息外化。对代理来说也是如此。用户必须尝试给代理许多不同的任务，以便了解代理的能力、如何清晰地描述目标、需要哪些技能文件以及哪些提示词最能引发所需的行为。随着模型的变化和改进，重新测试任务也很重要。提示词可能需要重新措辞，而曾经有用的护栏可能会限制更智能的模型寻求更具创造性的解决方案。

值得注意的是，我们发现最好的长期运行代理在有人类审阅之前，有许多不同的方法来验证其工作。代码当然有测试，但大多数其他工作也可以得到验证。例如，技术文档可以应用评分标准和风格指南。当人类设定标准并确保分配给代理的所有工作都可以接受审查时，质量就会保持高水平，且不会偏离最初的意图。另外，与人类一样，让一个代理负责完成任务，让另一个代理负责检查第一个代理的工作往往很有帮助，这通常被称为 “执行者-验证者”代理 harness。

在 Anthropic，与代理逐步建立信任表现为：

• 在开始时手动审查代理的工作，以审核质量、提供反馈并设计任务验证清单
• 告诉代理使用“验证者”代理来检查其工作，作为任务的一部分
• 将反思构建到循环中，并要求代理回顾自己的失误，以便工作随时间改进
• 跟踪每个代理在哪些类型的任务上获得了自主权，并在重复成功后按任务类型扩展范围

Anthropic 的一位工程负责人接手了一个有大量积压工作的新团队。为了控制局面，他邀请了几位人类同事和几个代理帮助他梳理积压任务并确定优先级。团队中的一组代理通读了积压中的所有条目，弄清了是否有人正在处理这些条目，并为任何无人负责的条目分配了复杂度评分。另一组则从列表中筛选出中低复杂度的条目，并创建了代码更改。一开始，人类审查了代理做出的每一个决定，并标记了任何需要人类输入的决定。然后，人类教会代理将这些决定直接呈现给人类，确保涉及艰难权衡的决策始终有人类参与。

每周，这位负责人和他的团队都会要求代理编制一份周报，其中包括“经验与失误”，以便代理跟踪错误并避免将来重蹈覆辙。随着时间的推移，这位负责人能够将越来越复杂的代码更改交给他的代理，并花更少的时间指导代理的日常任务。

一旦代理变得更加独立，负责人便指导它们将人类注意力视为稀缺资源：将问题批量整理以便一次性回答，重复关键上下文以快速让人类跟上进度，并限制每个人一次看到的事项数量。

帮助代理进行良好的沟通可以确保它们保持有用和高效。一些人在其团队中设置了仅负责决定如何批量和提升最重要通信给人类成员的代理。其他人则设置 guardra

Working with AI used to mean one person interfacing with a single chat window. Over time, AI has become increasingly capable at handling complex, long-running work, like coding, research, and financial analysis. With this, we’ve seen many new ways to use AI—from the terminal and IDE to spreadsheets and decks—but the work has still very much been a “single-player” experience: one human worked with one agent to accomplish individual tasks.

This is changing with the release of tools like Claude Tag. Now, humans and agents can work together in the same workspace, collaborating in service of goals shared by a team. Work now looks a lot more like a multiplayer game , with teams of humans setting the strategy, and Claude executing the work.

This involves some new ways of working. At Anthropic, we’ve been testing the technology required to make human-agent teams successful for the last several months. In this article, we explain what multiplayer agents are, and the lessons we’ve learned for building with them.

What are multiplayer agents?

“Multiplayer agents” is how we refer here to AI models that work with many different humans at the same time. Much like regular agents, they have their own memory and skills. But in other respects they're quite different. They have their own credentials and they live in places where work happens. At Anthropic, that's inside team collaboration tools like Slack.

Here’s an example of a human-agent team analyzing a dataset together in Slack:

For agents to productively participate in a team channel, they need specific capabilities:

These capabilities amount to the technical foundation required for an agent to participate productively across a team of many humans. However, making human-agent teams successful requires more than this: teams need specific ways of working and shared norms, too.

Lesson 1: Work in public and give agents broad context

Teams at Anthropic share information proactively and openly. This is especially true when agents are on the team, because agents build their understanding entirely from the text a team makes searchable: Slack, code, docs, and meeting notes. Private messages, hallway conversations, and restricted documents can’t provide agents with context. For an agent, if it’s not written down and accessible, it doesn’t exist.

Instead of deciding what information should be available to agents one doc or Slack channel at a time, we use clearly defined security boundaries that apply to entire Slack workspaces, as well as to meeting transcripts and doc libraries. Within the security boundary, context flows to every teammate—whether human or AI. Not only does this increase what agents and humans get access to, it also reduces confusion about what can be shared and with whom. Humans and agents alike find it difficult to navigate the soft boundaries of per-item sharing: should this channel be public or private? Can I share this doc with that person? Is this agent allowed to see that thread? A small number of clear, workspace-level boundaries removes decision fatigue from day-to-day work.

A high degree of transparency has a reward. For instance, agents that can read decisions from team meetings won't suggest tasks or projects that were deprioritized. Agents with access to product specs beyond their own team can recommend patterns that have succeeded for others. And because agents can read enormous volumes of text far faster than humans do, they routinely surface relevant work that humans would otherwise have missed. We lean on our agents heavily to stay informed and coordinated in a busy, fast-moving industry.

At Anthropic, working in public looks like:

• Choosing a handful of security boundaries at the company and creating workspaces and document sharing settings that match each security boundary
• Defaulting new communication channels to public within the organization, and ensuring decisions land in channels, docs, and meeting notes every time
• Writing artifacts and meeting notes so that agents can find them, since agents are now a primary consumer of team documentation
• Making sure AI has access to the right tools and information needed to get their job done

Defaulting information to be internally public can require cultural shifts. However, the difference between human-agent teams with context and those without is too stark to ignore.

Of course, some interactions are sensitive and will need to be private between a single human and AI. For those, with Claude Tag you can send @Claude a direct message, or you can use the existing Claude.ai and Claude Cowork applications. These tools give Claude access to private information via your personal MCP connectors, with the knowledge that your conversation and what you share with the agent will remain private.

Lesson 2: Every human and agent get a defined role with the right tools for the job

Human-agent teams share one roster, one set of artifacts, and one working space. Agents have their own credentials, skills, and tool access. Different agents also hold different roles: for instance, while one might own the data analysis for a project, another will hold and enforce the design standard, and a third will run research synthesis.

When a project kicks off, humans chat with the agents to figure out which roles to assign, and how the humans and agents will work together.

Once the jobs for humans and agents are clear, an agent might spin up other agents to make sure that specific tasks are handled by the agents with the right memory and appropriate access. Importantly, they need access to all the tools required to accomplish the job: one that handles data analysis might need access to BigQuery, and one that performs QA might need access to the Playwright MCP.

Clearly defined roles and responsibilities set human-agent teams up for success. Humans often work in the same threads the agents do, but they hold the roles only humans can hold. This ensures everything works together and human judgment is applied to the most important decisions. Without clear roles, people end up running fleets of personal AIs on the side, duplicating work and fracturing the team's context. Metrics tracking is a common case: a multiplayer agent can do the job once and let everyone see the same numbers.

At Anthropic, having clearly defined roles on human-agent teams looks like:

• An agreed-upon task set: the team's humans and its agents agree on who does what
• Humans and agents working in the same shared threads, so anyone can pick up where anyone left off
• Humans and agents that have access to the right tools to accomplish their respective jobs
• Descriptions of agents’ roles and scopes

Claude agents share the day-to-day maintenance of a codebase, triaging feedback, planning, writing code, reviewing changes, and reporting status. Each owns a clear task and works on its own schedule; people set the goals and review output.

An engineering team at Anthropic started creating rosters to help codify human and agent roles because it made driving their work much easier and more concrete. Some things that clicked for them early on:

• Specific roles also help humans easily track where responsibility for a task lies, whether that’s in individual tasks or an entire team’s set of responsibilities
• Writing skill files to define specific agents’ roles helps to make specialization easy, and allows people across the company to quickly stand up other agents of the same type
• The team adds new agents to focus on new areas when projects get more complex. For example, they added a release manager agent to deal with new software releases.

These methods let humans' mental model of a human-agent team scale as the number of agents grows.

Lesson 3: Set a north star to make agents more proactive

Although some agents at Anthropic simply complete assigned tasks, the most important ones proactively suggest new projects and workstreams. This often happens when a team that has already given its agents rich context and clear roles adds another guide: a north star.

North stars are ambitious, wide-reaching goals that help teams decide which tasks and workstreams are the right ones. At Anthropic, humans always set the north star, grounding it in the mission and goals of the business.

Once a north star is clearly articulated in writing, humans share it with the agents on their team. Then, importantly, humans choose which agents should proactively suggest new workstreams to help achieve this long-term goal. (It’s unlikely that every agent on the team will have the prerequisite skills and trust to proactively suggest work successfully.)

For example, an internal tools team with a north star to “make product onboarding more helpful” saw an agent proactively recommended copy revisions to the onboarding flow error messages. These changes measurably increased onboarding success the following week.

At Anthropic, setting a north star looks like:

• Having humans discuss, debate, and document an ambitious north star goal for their human-agent team—one that’s rooted in the company’s mission and business goals
• Sharing the north star with agents on the team and explicitly naming which agents can proactively recommend new workstreams
• Keeping high-fidelity human time protected on the calendar, with meetings now focused on the most important work

A clear north star gives agents a consistent direction to work toward and meaningful opportunities to proactively support a team’s work.

Lesson 4: Build trust over time

Teams at Anthropic grant agents autonomy in proportion to demonstrated reliability, then expand it deliberately. Engineers have successfully dispatched agents on their team to handle 500 bug fixes independently, but things certainly didn’t start off that way.

When a new human colleague joins the team, it takes time to assess their capabilities and develop strong working routines. It usually takes multiple feedback cycles to externalize all the tacit information about how tasks are best completed. The same is true for agents. Users have to experiment with giving agents many different tasks so they can learn what the agent is capable of, how to clearly describe the goal, what skill files it needs, and what prompts work best to elicit a desired behavior. It’s also important to retest tasks as models change and improve. Prompts may need re-wording and guardrails that used to be helpful may constrain a smarter model from pursuing more creative solutions.

Notably, we’ve found that the best long-running agents have many different ways to verify their work before a human looks at it. Code has tests, of course, but most other work can be verified as well. For example, technical docs can have rubrics and style guides applied to them. When humans set the bar and ensure all work assigned to an agent can be vetted, quality stays high and doesn’t drift from the original intention. Separately, as with humans, it often helps to give one agent the job of doing the task and another agent the job of checking the first agent’s work. This is often called the “Doer-Verifier” agent harness.

At Anthropic, building trust with agents over time looks like:

• Reviewing agent work manually in the beginning to vet quality, provide feedback, and design task verification checklists
• Telling the agent to use a “verifier” agent to check its work as part of the task
• Building reflection into the cycle and asking agents to review their own misses so work improves over time
• Tracking which kinds of tasks each agent has earned autonomy on and expanding scope per task type after repeated successes

One engineering leader at Anthropic took on a new team with a big backlog. To get a handle on it, he invited a few humans and a few agents to help him sort through the backlog and prioritize what was most important. One set of agents on the team read through all of the items in the backlog, figured out if anyone was working on the items, and assigned a complexity score to anything that was unowned. The other set read from the list, filtered to the medium and low complexity items, and created code changes. At the beginning, humans reviewed every decision made by an agent and marked any that required human input. Then the humans taught the agents to surface those decisions to humans directly, ensuring that decisions with hard tradeoffs always had a human in the loop.

Every week, the leader and his team asked the agents to compile a weekly report that included “lessons & missteps” so the agents would keep track of mistakes and avoid making them again in the future. Over time, the leader was able to give more and more complex code changes to his agents and spend less time guiding the agents’ day to day tasks.

And once the agents were more independent, the leader coached them to treat human attention as the scarce resource it is: to batch questions to be answered in a single pass, repeat key context to get a human up to speed quickly, and limit how many things each human sees at once.

Helping agents communicate well ensures that they remain helpful and effective. Some people have agents in their team with the sole role of deciding how to batch and elevate only the most important communication for human team members. Others set guardrails around how much work agents should do per day, so that humans are able to meaningfully engage with the work. Such guardrails ensure that humans maintain skills that are important to them, and that the number of items requiring human review stays sustainable.

Questions to ask

As you’re laying the foundation for your human-agent teams, consider the following questions:

1. Is all the information and access that agents and humans need both public and broadly searchable?
2. Can you write down your team's roster (humans and agents), and say what each member owns?
3. Does every human and agent on the team have access to the right tools to perform their job?
4. Do you have rubrics or tests for humans and agents to verify key work products?
5. Does your team have a clear north star that everyone can reference?

Moving forward

None of these patterns are new—at least not for humans. A strong north star, clear roles, strong documentation, a shared bar for quality, and room to learn from mistakes are the healthy team habits we’ve known for decades. Agents just make it even more important not to skip them.

The teams getting the most from their agents are the ones who are most intentional about applying these fundamentals.

Acknowledgements

This article was written by Kristen Swanson, a member of the Education team at Anthropic. She’d like to thank Matt Bell, Erik Olesund, Hasnain Lakhani, Shale Craig, Nolan Caudill, Mike Schiraldi, Aleks Todorova, and Molly Vorwerck for their contributions to this piece.

Start building multiplayer agents using agent teams in Claude Code or by using Claude Tag .