# Collaborative Intelligence
**作者**: Aatish Nayak
**日期**: 2021-08-02T14:10:57.000Z
**来源**: [https://x.com/nayakkayak/status/2009660549554913574](https://x.com/nayakkayak/status/2009660549554913574)
---
Today, AI works impressively for individuals but disappointingly for organizations. Closing that gap requires not just more context, but treating agents as social participants in the multiplayer systems they aim to disrupt.
To do this, many have concluded the answer is context graphs: a higher-order representation that stitches together decisions and records across systems. The assumption being that organizational context exists as a coherent structure just waiting to be discovered, like a fossil buried in shale.
In reality, most organizations do not operate from a centralized or intentionally designed logic. Context is not stored anywhere in full, not in databases, not in a clean outline, and certainly not in leadership’s head. It continuously emerges through interaction, with new context forming and decaying every day.
Only through treating context as the messy and unorganized social process it is can AI have any chance of learning it.
This requires agents to be embedded in the same collaboration primitives that humans use to work: email, messaging, calendars, browsers, documents, file systems, and more. They must observe how decisions unfold, learn which conflicts require escalation, and understand when consensus emerges versus when its imposed.
But this is only the beginning. The first wave of AI has mostly focused on individual intelligence. Each of us now have AI assistants that draft, analyze, and execute long range tasks with impressive reliability. For personal productivity, this represents a genuine step change. Inside organizations, however, the gains quickly plateau.
Why? Because in companies work happens between people through collaboration, negotiation, escalation, and shared judgment over time. One highly capable individual, human or AI, inside a misaligned organization does not meaningfully change outcomes. At best, it creates local efficiency. At worst, it entrenches unhealthy power dynamics.
To escape these constraints, we must realize that agents won’t map one to one with human roles. Instead, we must view agents as the renewable source of intelligence they are.
What changes when there is a tireless always-on team of infinite marketers, lawyers, engineers, analysts, and operators? How do organizations and teams reorganize to confront this reality? What novel interfaces must exist to allow each function to interact with a legion of agents?
This is the higher order opportunity for founders and builders. And a moment for deep reflection for enterprise leaders. Realizing this future will require messy experimentation, organizational restructuring, and a willingness to rethink how work gets done.
This is the only way individual intelligence evolves to collaborative intelligence.
Managing teams of agents will look more like Apollo Mission Control than a single command line: shared state and explicit escalation visible to all collaborators involved
## Social lessons from physics, telephones, and git
This evolution is not new. Humans faced it first before AI.
In Sapiens, Yuval Noah Harari argues that humans did not dominate the planet because of superior individual intelligence. Early humans were physically weaker than many animals and not meaningfully smarter than other hominins in isolation. What changed history was species-level cooperation.
Humans learned to collaborate in large groups by inventing shared stories like myths, laws, money, religions, and institutions, which allowed distributed actors to align behavior without central control.
Our history follows a consistent pattern. When new forms of technology emerge, they almost always outpace the collaborative systems needed to make them useful at scale.
Science offers the clearest illustration.
For most of human history, scientific knowledge did not compound reliably. Discoveries circulated only through private letters, books, or patronage networks. Errors persisted, insights were lost, and progress reset rather than accumulated. Almost like the goldfish mind of most AI products today.
The turning point was not a new theory, but the emergence of social systems for knowledge. In the seventeenth century, scientific societies formed, most notably the Royal Society of London. In 1665, Philosophical Transactions became the first scientific journal, establishing the norm that claims should be evaluated by a community rather than asserted by authority.
New claims were evaluated in relation to prior work, assumptions were challenged, and disagreements were resolved publicly within communities that transmitted judgment through apprenticeship as much as publication.
Astronomical observations of Jupiter and Saturn from multiple observers in London and Italy, published in Philosophical Transactions (1666) to reconcile differing measurements into a shared scientific record.
Another notable example was the Cavendish Laboratory at Cambridge. Under figures like J. J. Thomson and later Ernest Rutherford, generations of physicists learned how to reason about uncertainty, design experiments, interpret noisy data, and decide when evidence was sufficient. Knowledge began to compound because judgment became social across many actors.
Mass communication followed the same arc, but with an even clearer coordination bottleneck.
Early telephone systems connected callers point-to-point. If you knew exactly where the wire led, you could speak to the other end. As networks grew, this model collapsed. There was no way to dynamically route calls, prioritize connections, or manage contention. The technology worked, but the system did not.
Switchboards emerged as the social solution. Human operators sat at the center of early telephone networks, manually connecting calls by plugging cables into physical boards. They held context about who was calling, who was available, which connections were urgent, and how to resolve conflicts when multiple calls competed for the same line. Operators enforced norms, mediated errors, and adapted routing in real time.
Telephones scaled once calls were mediated through a shared, human-in-the-loop system rather than isolated endpoints.
The same pattern played out much more recently in software development.
Before Git, collaboration in software was fragile. Version-control systems like CVS and Subversion existed, but they were centralized and brittle. Working on the same codebase required serialized access, social coordination outside the system via email or meetings, and a high degree of trust. Conflicts were expensive, and history was shallow.
Linus Torvalds created Git in 2005 to manage the development of the Linux kernel after existing tools failed to scale. By making branching cheap and local, Git allowed many contributors to work independently without blocking one another. History was a first-class object preserving context about how and why the code evolved. Conflicts became explicit and resolvable rather than silent and destructive.
GitHub extended this further by layering explicit social coordination on top: pull requests, code review, issues, and visible discussion. Decision context accumulated around the code itself rather than disappearing into side channels.
Software development scaled because collaboration became first-class. We’re seeing this play out again as git/github have quickly become the default way AI coding agents are collaborating with humans.
The historical pattern in these and countless others is consistent. Individual capability appears first, and exponential productivity follows only once shared context and cooperative structures emerge.
## Reshaping orgs around collaboration units, instead of roles
To build these future cooperative structures, we must first envision what functions look like when they're no longer bound by human constraints: attention, bandwidth, specialization, and hierarchy.
Today's roles and headcount exist because work must be divided across people with limited time and context. Since agents don't share those constraints, collaborative AI will not simply assist existing roles, it will recombine them.
Instead of an organization designed around roles, it’ll instead be focused on collaboration units, unique to each function.
For example, in legal, the core unit is a shared position. Positions evolve through negotiation across associates, partners, and clients. Today, senior partners carry much of the connective context tracking precedent, posture, and unresolved risk across document turns and matters.
A collaborative legal agent absorbs much of this coordination. It tracks open issues across documents, surfaces conflicts in legal stance, and escalates judgment calls to the appropriate humans. Over time, the function reorganizes. The future legal deal team may consist of an army of agents doing mechanical drafting and internal information gathering, with the core decision-making, risk tolerance, and client relationship building done by a handful of senior partners.
In marketing, the challenge is narrative coherence across product marketing, growth, brand, and sales. Today this coherence is enforced through meetings, reviews, and informal influence.
A collaborative marketing agent participates directly in this work. It spans channels, remembers prior commitments, surfaces narrative drift, and escalates conflicts across teams. Roles shift away from channel ownership toward narrative stewardship and strategic intent. Agents provide continuity; humans provide taste and judgment.
In finance, work revolves around shared assumptions. Forecasts and budgets are collective agreements shaped through review, exception, and approval. Much of finance leadership’s value comes from arbitrating these assumptions across functions.
A collaborative finance agent tracks assumptions across scenarios, surfaces conflicts, routes exceptions, and preserves precedent. Over time, fewer roles focus on reconciliation and explanation, and more focus on risk framing, governance, and trust. Finance becomes less about producing numbers and more about overseeing decisions and balancing risk.
In product, the unit is the roadmap. What you prioritize, when you need it to happen, and how deep you need to invest in each item. Roadmaps evolve, constraints shift, and decisions are revisited as teams learn. Even today, most of the work of an EM, PM, or program manager is just coordinating all the resourcing, constraints, and execution timing of the roadmap.
A collaborative product agent operates as a shared participant in this process and facilitates these evolutions. It tracks unresolved tradeoffs, escalates tensions across roles, recommends resource reallocation, and preserves decision context as teams change. Humans spend less time re-establishing context and more time setting direction and understanding customer problems.
Multiplayer AI does not only make individual roles obsolete, but it redistributes coordination to these shared collaboration units, unique to each function. And eventually, these collaboration units should also intermingle with each other. For example, how should the roadmap change based on legal’s shared position on GDPR, or how should the marketing narrative change based on a planned roadmap release.
## Human escalation as an agent tool
This future won't emerge automatically. For founders and builders, the core challenge is treating collaboration as a first-class product primitive, not an afterthought.
The fastest path to multiplayer AI runs through the collaboration systems organizations already use. Email, messaging, browsers, and documents aren't legacy artifacts to avoid—they're the living infrastructure of work. Today, most entities we interact with on these platforms are still human. Why aren't we all pinging agents about how our day went? Or forwarding long email threads to agents for summaries?
These mediums encode how intent is expressed, how disagreement surfaces, how decisions escalate, and how accountability is recorded. Escalation is already built-in in the form of @ mentions, redlines, comments, suggested edits, notifications, and more.
One of the original drafts of the Declaration of Independence by Thomas Jefferson, with markups and redlines from others. Markups were the original form of escalation/feedback now codified into digital documents in the form of suggestions, comments, and more
In addition to gleaning valuable context, embedding agents into these surfaces also lowers the barrier to adoption. Instead of navigating to a URL or downloading an app, users simply do what they already do. Builders need to integrate aggressively with email clients, the Slack agent platform, and even browsers to access sources that aren't directly available.
(As an aside, this is the bull thesis on Atlassian buying The Browser Company and the Dia browser. Owning the browser layer for enterprise gives you instant access to the entire universe of SaaS tools without relying on messy integration battles.)
But these surfaces are only the starting point. As agents take on this more complex, long-horizon, and multi-team level work, they will need interfaces that go beyond our existing mediums.
Today, most AI interfaces focus on a single person working with a single AI, often only in a linear way. This quickly breaks when you assume you have an infinite factory of knowledge workers at your disposal, who constantly require your attention at different intervals. Moreover, these interfaces must define how agents should participate, escalate, and learn within functional systems and with their AI and human teammates. This is not only a product design problem, but an organizational design problem.
For each function and vertical, the interface will be different. Matter workspaces in legal where agents track open issues and escalate novel risks; assumption dashboards in finance where agents flag conflicts across forecasts; narrative control rooms in marketing where agents detect drift across campaigns & suggest improvements. These interfaces will not just display information, but encode the social structure of work itself with escalation and input built in.
For example, in legal an agent can sit inside a matter workspace: a dedicated interface that tracks open issues, pending decisions, and relevant precedents across all past matters. As the agent reviews a new inbound contract, it flags uncertainty and takes action based on role, risk, and precedent.
Routine deviations from standard language are resolved autonomously by referencing prior agreements. A novel but low-risk issue triggers an input request: the agent surfaces the clause in the workspace, attaches comparable precedents, and notifies an associate to confirm the approach. A material deviation on indemnity escalates directly to a partner, along with the client's risk posture, relevant precedents, and the decision to be made.
Conceptual design of a multi-player escalation interface with prior precedent, active agents at various stages of progress, and team members with pending decisions. Generalizable to many other functions
Critically, the interface doesn't just route decisions, but it encodes the collaboration itself. Over time, patterns accumulate. Certain classes of issues become autonomous. Others remain partner-owned but with clearer thresholds. The workspace becomes the living memory of the function: not just a place to view information, but a surface that governs how agents and humans collaborate.
The same structure generalizes across functions. In finance, assumptions escalate from analyst to controller to CFO based on materiality and precedent. In marketing, narrative conflicts escalate when brand commitments collide with growth experiments. In product, unresolved tradeoffs escalate when decisions cut across roadmap, reliability, and customer trust.
What changes with multiplayer AI is not that escalation suddenly appears. It is that escalation becomes explicit, legible, and persistent.
This is not a model problem, but a governance and people problem. AI must know who should have access to what, who has authority and who doesn’t, what context is shared vs local, and when to escalate appropriately. Builders must consider these features as foundations of their products instead of after thoughts. If collaboration is accidental, intelligence will remain local. If collaboration is architected, intelligence compounds.
This future will not arrive by accident.
## An organizational reckoning for enterprise leaders
So how does multiplayer AI get this valuable context? And how can enterprise leaders prepare?
Well, most organizations operate on a comforting fiction: that somewhere near the top, there exists a coherent understanding of how the system works. That strategy flows cleanly into execution. That coordination is designed rather than emergent.
In practice, this is rarely true.
In Moving off the Map, Ruthanne Huising shows what happens when employees attempt to map how their organization actually works. Even senior executives often believe someone, somewhere understands the system, who approves what, what data is stored where, and what escalation protocols need to happen. But the map shows that no such person exists.
Thank you to @emollick for first highlighting this before LLMs even appeared:
Organizations feel designed. In practice, they are organic. Work gets done through countless local decisions, informal escalations, side conversations, and judgment calls that aggregate over time.
This creates a second illusion: that context can be fully captured if only it is documented correctly. It cannot.
What leaders can do and must do, is articulate intent via shared stories. Mission, values, operating principles, escalation thresholds, and definitions of authority provide orientation. They define what the organization is trying to be, even if reality diverges in edge cases.
As the models get increasingly better, they will simple just need this intent in plain text, not in a artificially constructed context graph data structure. As long as the intent is clear, proactive escalation can take care of the rest.
For example, here’s an example intent for fraud review:
"Any transaction flagged for fraud review should be approved by the operations team if it follows our standard risk rubric and falls below $50K. Between $50K-$500K, escalate to the Risk Manager with context on why the case is borderline. Above $500K or involving a regulated entity, escalate to the Head of Compliance with full transaction history and comparable precedents."
This kind of intent doesn't prescribe every decision, but it provides clear orientation on authority boundaries, escalation thresholds, and information expectations. Everything else, how principles are applied under pressure, when rules bend, when exceptions are granted, which values win tradeoffs, is learned socially. No document captures it exhaustively, and no diagram ever will.
This distinction matters because it surfaces a deeper dynamic leaders must confront.
Much of modern management is coordination work. Middle managers, program leads, and functional heads spend a large portion of their time maintaining shared context, resolving ambiguity, aligning across silos, and escalating decisions. They act as human switchboards.
Multiplayer AI will reshape this layer directly.
As AI systems begin to track context, surface conflicts, and route decisions, the need for humans to act purely as connective tissue will change. Power that once came from asymmetric access to context becomes more visible. Organizational politics, often built on information control, will be at best upended, and at worst further entrenched.
Leaders who ignore this will struggle. Bloated organizations will be relics of the past, with headcount going towards compute vs compensation. Those who anticipate it can reshape leaner organizations around judgment, legitimacy, and direction.
Enterprise readiness for AI, then, is not about tooling alone. It is about whether leaders are willing to confront how coordination actually works inside their organizations and what will change when that coordination becomes shared with an infinite army of agents.
## 2026 is only the beginning
Individual AI agents are real progress, but the real frontier is cooperative, multiplayer intelligence. To truly unlock these capabilities will take a marriage of increasingly model capability, enterprise integrations, and organizations mandates to happen. In 2026, we will start to see the beginnings of this shift:
Shared workspaces will emerge as the first non-linear interfaces to work with AI. OpenAI, Anthropic, Google, Microsoft, and more will release shared team contexts where models retain state across users and tasks, not just isolated chat windows. As we’re already starting to see in coding IDEs, interfaces to manage legions of agents will emerge for all of knowledge work.
They will also release abstractions for enterprises and developers to build scoped memory, permissioned file system access, and more robust ways to incorporate tools into agents. Rather than static context graphs, we’ll see these active context surfaces that grow as agents participate in real work, escalating ambiguity to humans, capturing rationale, and making that precedent queryable and actionable.
Email, messaging, and systems of record will become execution layers. Google Workspace, Office 365, and others will continue embedding generative AI directly into mail and docs, with agents that draft, summarize, and route context-aware proposals across threads. Slack’s AI platforms will pull agents into channels where they can surface mismatches in narrative or decisions, suggest escalations, and track commitments over time. Microsoft will push Copilot deeper into Outlook and Teams, not just as an assistant but as a participant in group discussions, surfacing relevant context, unresolved issues, and next steps. Other SaaS companies like Salesforce, Workday, and Superhuman will continue to build on their agent platforms, encouraging startups to integrate to take advantage of their distribution.
Continual learning will emerge as a way for models to learn from collaboration. Model labs are already signaling that static training is not enough (see Nested Learning); that the next frontier is systems that learn from their environments, not just from curated datasets. But the most valuable environments are not synthetic benchmarks. They are organizations. Much like a new employee onboarding by participating in real work, future models will learn by operating inside live workflows. Continual learning, in this sense, will not look like autonomous self-improvement. It will look like socialization. And it will be the mechanism by which AI moves from capable participant to institutional actor over time.
Startups will build the missing vertical interfaces that incumbents overlook. We will see agents that live in vertical specific systems, and continuously reconcile changes, open questions, and historical judgments. This will be especially evident in domains like deal desks, underwriting, compliance, legal, and product planning, where “it depends” is the honest answer.
These shifts are not inevitable. They require intentional product choices: treating collaboration as a first-class primitive, anchoring in real workflows instead of side channels, and building governance that makes multiparty agent participation safe and legible.
One intelligent agent can help one person. Collaborative, cooperative, multiplayer intelligence is how AI helps organizations.
And that is the work ahead.
Thank you to @saranormous @ilyaf @saammotamedi @winstonweinberg @nikunj for feedback!
## 相关链接
- [Aatish Nayak](https://x.com/nayakkayak)
- [@nayakkayak](https://x.com/nayakkayak)
- [130K](https://x.com/nayakkayak/status/2009660549554913574/analytics)
- [Moving off the Map](https://gwern.net/doc/economics/2019-huising.pdf)
- [@emollick](https://x.com/@emollick)
- [Aug 2, 2021](https://x.com/emollick/status/1422198237344145410)
- [@saranormous](https://x.com/@saranormous)
- [@ilyaf](https://x.com/@ilyaf)
- [@saammotamedi](https://x.com/@saammotamedi)
- [@winstonweinberg](https://x.com/@winstonweinberg)
- [@nikunj](https://x.com/@nikunj)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [12:16 AM · Jan 10, 2026](https://x.com/nayakkayak/status/2009660549554913574)
---
*导出时间: 2026/1/14 19:41:55*
---
## 中文翻译
# 协作智能
**作者**： Aatish Nayak
**日期**: 2021-08-02T14:10:57.000Z
**来源**： [https://x.com/nayakkayak/status/2009660549554913574](https://x.com/nayakkayak/status/2009660549554913574)
---
今天，AI 在个人层面的表现令人印象深刻，但在组织层面却令人失望。弥合这一差距不仅需要更多的上下文，还需要将智能体视为其试图变革的多方系统中的社会参与者。
为此，许多人得出的结论是答案在于上下文图谱：一种高阶表示，将跨系统的决策和记录拼接在一起。其假设是，组织上下文以一种连贯的结构存在，只等着被发现，就像埋在页岩中的化石一样。
实际上，大多数组织并非基于集中式或有意设计的逻辑运作。上下文并没有在任何地方被完整存储——不在数据库中，不在清晰的提纲中，当然也不在领导者的头脑中。它通过互动不断涌现，每天都有新的上下文形成和消亡。
只有把上下文视为它本来的样子——一个混乱而无序的社会过程——AI 才有机会学习到它。
这要求智能体嵌入到人类用于工作的相同协作原语中：电子邮件、即时通讯、日历、浏览器、文档、文件系统等等。它们必须观察决策是如何展开的，学习哪些冲突需要上报，理解共识是何时自然形成的，以及何时是被强行施加的。
但这仅仅是开始。第一波 AI 主要聚焦于个体智能。我们每个人都已拥有 AI 助手，它们能够起草、分析并以令人惊叹的可靠性执行长程任务。对于个人生产力而言，这是一个真正的飞跃。然而在组织内部，收益却很快趋于平缓。
为什么？因为在公司中，工作是通过人与人之间的协作、谈判、升级和长期的共同判断来完成的。在一个失调的组织中，一个能力再强的个体——无论是人类还是 AI——都无法真正改变结果。往好处说，它只能带来局部的效率提升；往坏处说，它会固化不健康的权力关系。
要摆脱这些束缚，我们必须认识到智能体不会与人类角色一一对应。相反，我们必须将智能体视为它们本质上的样子——一种可再生的智能源泉。
当拥有一支不知疲倦、全天候在线的团队，由无限的营销人员、律师、工程师、分析师和运营人员组成时，会发生什么变化？组织和团队将如何重组以应对这一现实？需要哪些新颖的界面，才能让每个职能与一支智能体军团互动？
这是创始人和建设者面临的高阶机遇。也是企业领导者进行深度反思的时刻。实现这一未来将需要混乱的实验、组织重构，以及重新思考工作完成方式的意愿。
这是个体智能演化为协作智能的唯一途径。
管理智能体团队将更像阿波罗任务控制中心，而不是单一命令行：共享状态和明确的升级机制对所有参与的协作者可见
## 来自物理学、电话和 git 的社会启示
这一演化并不新鲜。人类在 AI 之前就早已面对过它。
在《人类简史》中，尤瓦尔·赫拉利指出，人类主宰地球并非因为个体智能的卓越。早期人类在体能上弱于许多动物，单独来看也不比其他人属物种聪明多少。真正改变历史的是物种层面的合作。
人类通过发明共享的故事——神话、法律、货币、宗教和制度——学会了在大群体中协作，这使得分散的行动者能够在没有中央控制的情况下协调行为。
我们的历史遵循着一个一致的模式。当新形式的技术出现时，它们几乎总是超前于使其大规模发挥作用的协作系统。
科学提供了最清晰的例证。
在人类历史的大部分时间里，科学知识并不可靠地累积。发现仅通过私人信件、书籍或赞助人网络传播。错误持续存在，洞见丢失，进展被重置而非累积。几乎就像今天大多数 AI 产品的金鱼式记忆。
转折点不是某个新理论，而是知识的社会系统的出现。十七世纪，科学学会成立，其中最著名的是伦敦皇家学会。1665 年，《哲学汇刊》成为第一份科学期刊，确立了一项规范：主张应由共同体来评估，而非由权威来断言。
新的主张要参照先前的工作来评估，假设会受到挑战，分歧在共同体内部公开解决——这些共同体通过师徒传承和出版物同样地传递着判断力。
伦敦和意大利多位观测者对木星和土星的天文观测，发表于《哲学汇刊》（1666 年），旨在将不同的测量结果调和为共享的科学记录。
另一个著名的例子是剑桥大学的卡文迪许实验室。在 J. J. 汤姆逊以及后来的欧内斯特·卢瑟福等人物的带领下，一代代物理学家学会了如何推理不确定性、设计实验、解读含噪数据，以及判断证据何时足够。知识开始复利式增长，因为判断力在众多行动者之间变成了社会性的。
大规模通信也遵循了同样的轨迹，但协调瓶颈更为明显。
早期的电话系统将呼叫者点对点连接。如果你确切知道电线通向哪里，你就可以与另一端通话。随着网络的发展，这种模式崩溃了。没有办法动态路由呼叫、优先处理连接或管理竞争。技术可行，但系统不行。
电话交换机应运而生，成为社会层面的解决方案。人工操作员坐在早期电话网络的中心，通过将电缆插入物理配线板来手动接通电话。他们掌握着各种上下文信息：谁在打电话、谁有空闲、哪些通话是紧急的，以及当多个通话争用同一条线路时如何化解冲突。操作员执行规范、调解错误，并实时调整路由。
当通话通过一个共享的、人在回路中的系统而非孤立的端点来协调时，电话才得以规模化。
同样的模式最近在软件开发领域再次上演。
在 Git 出现之前，软件协作是脆弱的。CVS 和 Subversion 等版本控制系统虽然存在，但它们是中心化的且脆弱。在同一个代码库上工作需要串行访问、通过邮件或会议在系统之外进行社会协调，以及高度信任。冲突代价高昂，历史记录也浅薄。
2005 年，Linus Torvalds 在现有工具无法扩展之后，创建了 Git 来管理 Linux 内核的开发。通过让分支创建变得廉价且本地化，Git 允许众多贡献者独立工作而互不阻塞。历史记录成为了一等公民对象，保存着代码如何演化以及为何演化的上下文。冲突变得明确且可解决，而不是隐蔽且具有破坏性。
GitHub 在此基础上进一步扩展，叠加了显式的社会协调层：拉取请求、代码审查、issue 和可见的讨论。决策上下文围绕代码本身积累，而不是消失在旁路渠道中。
软件开发得以规模化，是因为协作成为了一等公民。我们正再次见证这一过程——git/github 已迅速成为 AI 编程智能体与人类协作的默认方式。
这些案例以及其他无数案例中的历史模式是一致的：个体能力率先出现，而只有当共享上下文与合作结构形成后，生产力的指数级增长才会随之而来。
## 围绕协作单元而非角色重塑组织
要构建这些未来的合作结构，我们必须首先设想：当职能不再受人类约束——注意力、带宽、专业化和层级——的束缚时，它们会是什么样子。
今天的角色和编制之所以存在，是因为工作必须在时间和上下文有限的人群之间划分。而智能体不受这些约束，因此协作式 AI 不会简单地辅助现有角色，而是会重组它们。
组织将不再是围绕角色来设计，而是围绕协作单元来构建，每个职能都有其独特的协作单元。
例如，在法律领域，核心单元是共同的立场。立场通过合伙人、律师和客户之间的协商而演变。如今，资深合伙人承担着大量关联性上下文的维护工作——在文档轮次和事务之间追踪先例、立场姿态和未解决的风险。
协作式法律智能体可以吸收这些协调工作的很大一部分。它跨文档追踪未决问题，发现法律立场上的冲突，并将需要判断的事项升级给合适的人。久而久之，该职能将重新组织。未来的法律交易团队可能由一支智能体大军组成，负责机械性的起草和内部信息收集，而核心的决策、风险容忍度和客户关系维护则由少数几位资深合伙人完成。
在市场营销领域，挑战是在产品营销、增长、品牌和销售之间保持叙事的一致性。如今，这种一致性是通过会议、评审和非正式影响力来维持的。
协作式营销智能体直接参与到这项工作中。它跨越各个渠道，记住先前的承诺，发现叙事的漂移，并在团队之间将冲突升级。角色从渠道所有权转向叙事管理和战略意图。智能体提供连续性；人类提供品味和判断。
在财务领域，工作围绕共享假设展开。预测和预算是通过评审、例外处理和审批形成的集体共识。财务领导层的很大一部分价值来自于跨职能仲裁这些假设。
协作式财务智能体跨场景追踪假设，发现冲突，路由异常情况，并保留先例。久而久之，更少的角色专注于对账和解释，更多的角色专注于风险框架构建、治理和信任。财务将不再是产出数字，而是更多地监督决策和平衡风险。
在产品领域，单元是路线图。你要确定优先级、时间要求，以及对每个事项需要投入多深。路线图在演变，约束在变化，随着团队的学习，决策也会被重新审视。即使在今天，EM、PM 或项目协调人（program manager）的大部分工作也只是协调路线图的资源配置、约束和执行时间。
协作式产品智能体作为这一过程中的共享参与者运作，并促进这些演变。它追踪未决的权衡，跨角色升级紧张关系，推荐资源重新分配，并在团队变动时保留决策上下文。人类花更少的时间重建上下文，而花更多的时间设定方向和理解客户问题。
多人协作式 AI 不仅会让个体角色过时，还会将协调工作重新分配到这些共享的协作单元上，每个职能都有其独特的单元。最终，这些协作单元之间也应该相互交融。例如，路线图应如何根据法律部门在 GDPR 上的共同立场做出调整，或者营销叙事应如何根据计划中的路线图发布而做出改变。
## 作为智能体工具的人工升级
这个未来不会自动到来。对于创始人和建设者来说，核心挑战是把协作当作一等公民的产品原语，而不是事后补充。
通往多人AI的最快路径，就贯穿于组织已经在使用的协作系统之中。电子邮件、即时通讯、浏览器和文档并不是需要避开的遗留产物——它们正是工作运转的活基础设施。如今，我们在这些平台上互动的大多数对象仍然是人类。为什么我们不向agent发送消息问问自己今天过得怎么样？或者把冗长的邮件串转发给agent让其总结？
这些媒介编码了意图如何表达、分歧如何浮现、决策如何升级，以及责任如何记录。升级机制早已内置于@提及、修订标记、评论、建议修改、通知等功能之中。
托马斯·杰斐逊《独立宣言》的原始草稿之一，带有他人的批注和修订标记。批注是最初的升级/反馈形式，如今以建议、评论等形式被固化到数字文档中
除了获取有价值的上下文之外，将agent嵌入这些界面还能降低采用门槛。用户无需访问某个URL或下载应用，只需做他们已经在做的事情。开发者需要积极与邮件客户端、Slack agent平台，甚至浏览器进行集成，以获取那些无法直接获得的资源。
（顺便一提，这正是看多Atlassian收购The Browser Company和Dia浏览器这一交易的理由。掌控企业浏览器层，就等于无需依赖繁琐的集成之争，即可即时访问整个SaaS工具宇宙。）
但这些界面只是起点。随着agent承担起更复杂、更长期、跨多团队的工作，它们将需要超越现有媒介的界面。
如今，大多数AI界面聚焦于一个人与一个AI协作，且往往只是线性方式。当你设想拥有一个无限的知识工作者工厂、它们在不同时间点不断需要你的关注时，这种方式很快就会失效。此外，这些界面还必须定义agent应如何在职能系统中、与其AI和人类队友之间参与、升级和学习。这不仅是产品设计问题，更是组织设计问题。
对每个职能和垂直领域而言，界面都将有所不同。在法律领域是matter工作区，agent在其中追踪未决问题并升级新出现的风险；在财务领域是假设看板，agent在其中标记跨预测的冲突；在营销领域是叙事控制室，agent在其中检测各campaign之间的偏差并提出改进建议。这些界面不仅会展示信息，还会编码工作本身的社会结构，将升级和输入机制内建其中。
例如，在法律领域，agent可以置身于一个matter工作区之中：这是一个专门的界面，追踪所有过往matter中未决的问题、待定的决策以及相关先例。当agent审查一份新收到的合同时，它会标记不确定性，并根据角色、风险和先例采取行动。
对标准条款的常规偏离，agent会参考过往协议自主解决。新颖但低风险的问题会触发一次输入请求：agent在工作区中呈现该条款，附上可比先例，并通知一位律师确认处理方式。而赔偿条款的重大偏离则会直接升级给合伙人，同时附上客户的风险偏好、相关先例以及待做出的决策。
多人升级界面的概念设计，包含过往先例、处于不同进展阶段的活跃agent，以及有待决决策的团队成员。可推广到许多其他职能领域
至关重要的是，这个界面不仅仅路由决策，它还编码了协作本身。随着时间推移，模式不断积累。某些类别的问题变为自主处理。另一些仍由合伙人负责，但阈值更加清晰。工作区成为该职能的活性记忆：不仅是查看信息的场所，更是规范agent与人类如何协作的界面。
同样的结构可以推广到各个职能。在财务领域，假设会根据重要性和先例，从分析师升级到财务主管再到CFO。在营销领域，当品牌承诺与增长实验相冲突时，叙事冲突就会升级。在产品领域，当决策横跨路线图、可靠性和客户信任时，未解决的权衡就会升级。
多人AI带来的变化，并不是升级机制突然出现。而是升级机制变得明确、可读且持久。
这不是模型问题，而是治理和人的问题。AI必须知道谁应该有权访问什么，谁有权限谁没有，哪些上下文是共享的、哪些是本地的，以及何时该恰当地升级。开发者必须将这些特性视为产品的基础，而不是事后才想起的补充。如果协作是偶然发生的，智能将始终局限于局部。如果协作是被精心架构的，智能就会复利增长。
这个未来不会偶然到来。
## 企业领导者的组织性清算
那么，多人AI如何获得这些宝贵的上下文？企业领导者又该如何准备？
大多数组织都建立在一种令人心安的虚构之上：在接近顶层的地方，存在着对系统如何运作的连贯理解。战略可以顺畅地传导为执行。协调是设计出来的，而非自然涌现的。
在实践中，这很少是事实。
在《Moving off the Map》中，Ruthanne Huising展示了当员工试图绘制组织实际运作方式时会发生什么。即便是高管，也常常认为某处的某个人理解整个系统——谁批准什么、什么数据存储在哪里、需要执行什么样的升级流程。但地图显示，这样的人并不存在。
感谢 @emollick 在LLM尚未出现时就率先指出了这一点：
组织感觉上是经过设计的，但实际上它们是有机的。工作是通过无数局部的决策、非正式的升级、私下的对话和判断取舍完成的，这些随着时间推移不断累积。
这造成了第二种幻觉：认为只要正确地记录下来，上下文就可以被完全捕捉。它做不到。
领导者能够做、也必须做的，是通过共享的故事来阐明意图。使命、价值观、运营原则、升级阈值和权限定义提供了方向。它们定义了组织想要成为什么，即使现实在边缘情况下有所偏离。
随着模型越来越好，它们只需要以纯文本形式获得这种意图，而不需要人工构建的上下文图数据结构。只要意图清晰，主动升级就能处理其余的事情。
例如，以下是欺诈审查意图的一个示例：
"任何被标记为欺诈审查的交易，如果符合我们的标准风险规则且金额低于 5 万美元，应由运营团队批准。在 5 万至 50 万美元之间，应升级给风险经理，并提供该案例处于边界情况的原因背景。超过 50 万美元或涉及受监管实体，应升级给合规主管，并提供完整的交易历史和可比的先例。"
这种意图并不规定每一个决策，但它为权限边界、升级阈值和信息预期提供了清晰的方向。其余的一切——原则在压力下如何应用、规则何时可以变通、何时给予例外、哪些价值观在权衡中胜出——都是通过社会化习得的。没有任何文档能详尽地记录它，任何图表也永远无法做到。
这一区别很重要，因为它揭示了领导者必须面对的一个更深层的动态。
现代管理的大部分工作是协调工作。中层管理者、项目负责人和职能主管将大量时间用于维护共享上下文、消除歧义、跨部门对齐以及升级决策。他们就像人工交换机。
多智能体 AI 将直接重塑这一层。
随着 AI 系统开始追踪上下文、发现冲突并路由决策，人类纯粹充当连接组织的需要将发生变化。曾经来自对上下文的不对称获取的权力将变得更加透明。往往建立在信息控制之上的组织政治，往好了说将被颠覆，往坏了说将被进一步固化。
忽视这一点的领导者将会陷入困境。臃肿的组织将成为过去的遗物，人员编制将转向算力而非薪酬。而那些能够预见到这一点的人，可以围绕判断力、合法性和方向感重塑更精简的组织。
因此，企业的 AI 就绪不仅仅是关于工具。它关乎领导者是否愿意直面组织内部协调的实际运作方式，以及当这种协调与一支无限的智能体大军共享时，将会发生什么变化。
## 2026 年只是开始
单个 AI 智能体是真正的进步，但真正的前沿是协作式的、多玩家的智能。要真正释放这些能力，需要日益增强的模型能力、企业集成和组织授权三者的结合。在 2026 年，我们将开始看到这一转变的开端：
共享工作区将成为与 AI 协作的首批非线性界面。OpenAI、Anthropic、Google、微软等将推出共享的团队上下文，模型可以在用户和任务之间保持状态，而不仅仅是孤立的聊天窗口。正如我们在编程 IDE 中已经开始看到的那样，管理智能体军团的各种界面将出现在所有知识工作中。
它们还将为企业和开发者发布各种抽象，用于构建范围化的记忆、有权限的文件系统访问，以及将工具整合到智能体中的更健壮的方式。我们将看到的不是静态的上下文图，而是这些活跃的上下文界面，它们随着智能体参与实际工作而增长，将模糊之处升级给人类、捕捉决策理由，并使这些先例可查询、可执行。
电子邮件、消息传递和记录系统将成为执行层。Google Workspace、Office 365 等将继续把生成式 AI 直接嵌入到邮件和文档中，由智能体在对话线程中起草、总结并路由具有上下文感知的提案。Slack 的 AI 平台将把智能体引入频道，在其中它们可以发现叙述或决策中的不一致之处，建议升级，并随时间追踪承诺事项。微软将把 Copilot 更深入地推进到 Outlook 和 Teams 中，不仅是作为助手，而是作为小组讨论的参与者，呈现相关的上下文、未解决的问题和后续步骤。Salesforce、Workday 和 Superhuman 等其他 SaaS 公司将继续在其智能体平台上进行构建，鼓励初创公司进行集成以利用其分发渠道。
持续学习将作为一种让模型从协作中学习的方式出现。各大模型实验室已经发出信号，表明静态训练是不够的（参见嵌套学习）；下一个前沿是能够从环境中学习的系统，而不仅仅是从人工整理的数据集中学习。但最有价值的环境不是合成基准测试，而是组织。就像新员工通过参与实际工作来入职一样，未来的模型将通过在真实的工作流程中运作来学习。从这个意义上说，持续学习不会表现为自主的自我改进，而是表现为社会化。它将成为 AI 随时间从有能力的参与者转变为机构行为者的机制。
初创公司将构建被现有巨头忽视的缺失的垂直界面。我们将看到生活在垂直领域特定系统中的智能体，持续协调变更、悬而未决的问题和历史判断。这在交易谈判台、承保、合规、法律和产品规划等领域尤为明显，因为在这些领域，“视情况而定”才是诚实的答案。
这些转变并非注定会发生。它们需要有意识的产品选择：将协作视为一等原语，扎根于真实的工作流程而非旁支渠道，并构建能够让多方智能体参与变得安全且清晰可见的治理机制。
一个智能体可以帮到一个人。协作的、合作的、多人参与的智能，才是 AI 帮助组织的方式。
而这正是未来的工作。
感谢 @saranormous @ilyaf @saammotamedi @winstonweinberg @nikunj 提供反馈！
## 相关链接
- [Aatish Nayak](https://x.com/nayakkayak)
- [@nayakkayak](https://x.com/nayakkayak)
- [130K](https://x.com/nayakkayak/status/2009660549554913574/analytics)
- [Moving off the Map](https://gwern.net/doc/economics/2019-huising.pdf)
- [@emollick](https://x.com/@emollick)
- [2021年8月2日](https://x.com/emollick/status/1422198237344145410)
- [@saranormous](https://x.com/@saranormous)
- [@ilyaf](https://x.com/@ilyaf)
- [@saammotamedi](https://x.com/@saammotamedi)
- [@winstonweinberg](https://x.com/@winstonweinberg)
- [@nikunj](https://x.com/@nikunj)
- [升级到高级版](https://x.com/i/premium_sign_up)
- [2026年1月10日 上午12:16](https://x.com/nayakkayak/status/2009660549554913574)
---
*导出时间: 2026/1/14 19:41:55*
