智东西（公众号：zhidxcom）
作者 | ZeR0
编辑 | 漠影
智东西11月4日报道，今日，2026世界互联网大会乌镇峰会浙江分论坛举行。会上，网易智企副总经理段毓铮发表《智能体：让美好发生在每一个现场》主题演讲，发布网易智企企业级AI Agent平台“帝王蟹”的三项升级：首发硬件产品智能工牌、推出知识引擎“万典”与Agent Team两项新能力。
三项升级分别聚焦现场信息接入、企业知识理解与复杂任务协作，进一步串联AI进入真实业务所需要的信息、知识与执行环节。
▲网易智企副总经理段毓铮
段毓铮谈道，企业对AI的关注正在从“模型有多强”“用什么Agent”，逐渐转向“AI能不能真正进入业务、持续创造价值”，这意味着企业AI正从技术验证走向生产力验证。
随着模型能力逐步普及，同一模型进入不同企业后，实际业务效果的差异开始受到更多关注。段毓铮认为，未来真正拉开企业差距的，不只是采用了什么模型，而是AI能否理解一家企业自己的业务、组织、知识、经验、流程、规则和做事方式。
他将这些长期沉淀在企业经营和协作过程中的信息、知识、规则、流程和实践经验概括为“企业上下文”。
与模型能力相比，一家企业长期形成的业务逻辑和实践经验更难被简单复制。随着AI更深入地进入业务，这些企业特有的信息和做事方式，也会越来越直接地影响AI的实际工作效果。
围绕如何让企业上下文进入具体任务，帝王蟹此次带来了三项升级。
一、智能工牌，将线下业务信息接入AI系统
企业业务并不只发生在线上系统中。客户拜访、门店巡检、现场服务、内部沟通等场景每天都会产生大量一线信息，但这些内容往往停留在线下，难以进入企业现有的信息系统，也很难被AI进一步理解和利用。
此次首发的硬件产品智能工牌，主要用于将这部分现场信息接入AI系统。
在客户拜访等场景中，线下沟通内容可以通过智能工牌进入帝王蟹，由平台完成后续分析和处理。设备内置 SIM 卡，可直接连接云端，同时支持物理隐私开关、录音状态提示和录音时段管理，让现场信息的进入可见、可管。
由此，AI可以处理的信息不再局限于文档、CRM等线上系统，而是进一步延伸至真实业务现场。客户需求、一线问题和现场反馈，也可以进入后续业务分析和任务执行。
例如，一项客户需求可能先出现在拜访录音中，随后进入CRM、邮件或工单，并进一步影响方案、报价和交付。
二、“万典”，让企业知识可找、可信、可用
现场信息进入系统之后，企业还面临另一个问题：完成一项任务所需要的信息，往往分散在不同系统中，并且持续变化。
文档、邮件、工单、CRM、数据库、第三方知识服务，以及客户拜访、会议和现场沟通，都可能包含重要依据。面对复杂业务任务，AI不仅需要找到相关内容，还要理解信息之间的关联，判断哪些事实可靠、哪些已经变化。
知识引擎“万典”要解决的，是让Agent面对复杂任务时，能够持续理解和整合企业中不断变化的信息，并形成更完整、可靠的任务依据。
它可以围绕具体任务，动态组织文档、邮件、工单、CRM、数据库及现场沟通等不同来源、不同时间的信息，持续查找、补证和核验，并识别不同信息和事件之间的逻辑与时序关系。
比如，一项客户需求可能先出现在拜访录音中，随后进入CRM、邮件或工单，并进一步影响方案、报价和交付。
“万典”不仅要找到相关信息，还要判断哪些是最新事实、哪些能够相互印证、还缺哪些依据，并形成可供Agent后续判断和执行使用的任务依据。
证据不足时继续查找，无法确认时如实说明，相关结论还可以追溯至原文、具体页码或录音中的对应时间点。
获得足够信息后，Agent可继续调用业务工具，将知识和分析结果带入后续任务执行；同一套知识和权限体系，也可以被不同Agent复用。
三、Agent Team，让多个Agent协同完成复杂任务
针对跨角色、跨步骤的复杂任务，帝王蟹推出了Agent Team。
企业实际业务中，一项任务往往需要多个专业角色共同参与。以合同审查为例，一份合同可能同时涉及商务、财务、风控和法务等多个部门，不同角色关注的问题并不相同。
在Agent Team中，一个协调Agent负责拆解任务、分配工作和汇总结果，不同专业角色的Agent分别处理相应任务，并可围绕同一问题进行讨论、质疑和校验。
对于跨部门、多步骤的流程，不同Agent还可以根据任务进展相互沟通和交接。可并行的任务同步处理，存在前后依赖的环节则按顺序推进。
段毓铮称，多Agent协作一方面可以让不同专业视角共同参与复杂问题的判断，在相互讨论和校验中提升决策质量；另一方面，也可以推动原本依赖人工协调的长链路流程持续推进，提高流程的自动化和自主执行能力。
在关键判断和重要决策环节，仍由人保留授权和最终决策。
四、三项升级背后，是一家企业自己的“上下文”
从现场信息进入系统，到多源企业知识被理解和使用，再到多个Agent协同完成复杂任务，三项升级串联了AI进入真实业务所需要的信息、知识与执行链路，让企业上下文能够更完整地进入具体任务的判断和执行过程。
随着这条链路持续运行，企业长期积累的知识、流程、规则、经验和协作方式，也会越来越完整地进入AI的工作过程，共同构成一家企业自己的“企业上下文”。
网易智企长期服务企业通信与协作场景，即时通信与音视频服务月均消息量超过3400亿条，接入企业开发者235万。
长期参与企业内部以及企业与客户、供应商之间的信息流转，也使其持续接触到企业信息如何产生、流动和沉淀，以及不同角色之间如何协同。
在企业AI的落地路径上，段毓铮认为，不必一开始就追求“大而全”，更现实的方式是从具体场景和任务切入，把高价值问题做深，再逐步向前后环节延伸，连接成完整业务链路，让AI从单点提效进一步走向经营提效。
他将这一过程概括为：场景落地、闭环打通、经营提效。
企业可以从销售、研发、生产、服务等具体任务入手，再向上下游延伸，形成“线索到回款”“需求到交付”“问题到解决”“战略到执行”等业务闭环。
结语：企业AI从完成任务，走向参与企业判断
随着AI沿业务链路进入更多环节，其能够接触到的一线信息、企业知识和业务过程也会逐渐完整。
在此基础上，多Agent还可以从不同专业视角参与讨论和校验，AI在企业分析和判断中的参与度也会进一步提高。
更长远看，帝王蟹的目标是进一步成为企业的“决策引擎”。
段毓铮认为，这并不意味着让AI替代人作出最终决定，而是让更多真实业务信息和专业视角进入判断过程，帮助企业做出更快、更高质量的决策，人在关键环节保留授权和最终决策。
从把事情做完，到帮助企业把事情想得更清楚、判断得更准确，这将是企业AI继续向前的重要一步。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
Author ZeR0
Edit |
On 4 November, Jijiang reported that today, the Zhejiang sub-forum of the U-town Summit of the 2026 World Internet Congress was held. At the conference, the Deputy Managing Director of the Internet-Intelligence Enterprise published the keynote address AI agent: Making Good happen on Every Live, and three upgrades of the publish Internet-Intelligent Enterprise-level AI Agent platform, the King's Crab: the launch of the hardware product smart card, the launch of the Knowledge Engine “Wanks” and Agent Team.
The three upgrades focus on on on-site information access, enterprise knowledge understanding and collaboration with complex missions, further linking AI to the information, knowledge and implementation links required for real operations.
Deputy Director-General of the Network
It was argued that business concerns about AI were moving from “how strong the model” to “what to use” to “how real AI can enter business and create value on a continuous basis”, meaning that it was moving from technical to productivity validation.
As model capabilities become more widespread and the same model enters different enterprises, the difference in actual operational effectiveness begins to gain more attention. It is argued that what really opens the business gap in the future is not just the model used, but whether AI understands an enterprise ' s own business, organization, knowledge, experience, processes, rules and ways of doing things.
He summarized these long-standing information, knowledge, rules, processes and practical experiences in the course of business operations and collaboration as “enterprise context”.
The long-standing business logic and practical experience of an enterprise is more difficult to easily replicate than modelling capacity. As AI moves deeper into business, the information and ways in which these enterprises conduct their business will increasingly directly influence the actual performance of AI.
The Krabs brought three upgrades this time around how to get the company context into a specific mission.
I. Smart cards to access the AI system for offline business information
Businesses do not only operate online. Client visits, door-to-door inspections, on-site services, and internal communication generate a large amount of information on a daily basis, but these content often remain on-line, difficult to access to existing enterprise information systems and to be further understood and used by AI.
This hardware product smart card was launched, mainly to access this part of the field information into AI.
In scenarios such as customer visits, belowline communication content can enter the King ' s Krabs through smart work cards, with the platform completing follow-up analysis and processing. The equipment is built into a SIM card, which connects directly to the cloud, while supporting physical privacy switches, audio state alerts and time management for recording, so that information from the site is visible and manageable.
As a result, the information that AI can process is no longer limited to online systems such as documents, CRMs, but extends further to real business sites. Client needs, front-line issues and on-site feedback are also available for subsequent business analysis and task execution.
For example, a customer demand may appear first in a call recording and then enter a CRM, mail or worksheet, further affecting the programme, offer and delivery.
II. “All-round” to make business knowledge searchable, credible and usable
Once on-site information enters the system, the enterprise faces another problem: the information required to perform a mission is often dispersed across different systems and constantly changing.
Documents, mail, worksheets, CRMs, databases, third-party knowledge services, as well as customer visits, meetings and on-site communication may all contain important grounds. Faced with complex operational tasks, AI needs not only to find relevant elements but also to understand the links between information and determine what is reliable and what has changed.
What the knowledge engine, the "one-size-fits-all", has to do is to enable Agent to continuously understand and integrate changing information in an enterprise when faced with complex tasks and to form a more complete and reliable mandate.
It can organize information from different sources, at different times, around specific tasks, such as documents, mail, worksheets, CRMs, databases and on-site communications, continuously search, back up and verify, and identify the logical and time-series relationships between different information and events.
For example, a customer demand may appear first in a visit recording and then enter a CRM, mail or worksheet, further affecting the programme, offer and delivery.
Not only is it necessary to find relevant information, but it is also necessary to determine what is the latest facts, what is mutually verifiable and what is missing, and to form the basis of the mandate for Agent's follow-up judgement and implementation.
If there is insufficient evidence to continue the search, and if this is not confirmed, the conclusions can also be traced back to the original, the specific page number or the corresponding point in the recording.
Once sufficient information is available, Agent can continue to use operational tools to bring knowledge and analysis to follow-up missions; the same body of knowledge and permission systems can also be reused by different Agents.
iii. Agent Team, with multiple Agents working together on complex tasks
In response to complex tasks across roles and steps, the King Krabs launched Agent Team.
In the actual business of an enterprise, a task often involves multiple professional actors. In the case of contract reviews, for example, a contract may involve multiple sectors, such as commerce, finance, control and law, with different players having different concerns.
In Agent Team, a coordinator of Agent was responsible for dismantling tasks, assigning tasks and consolidating results, and Agent, from different professional actors, dealt with the respective tasks and could discuss, challenge and validate the same issue.
For cross-sectoral, multi-step processes, different Agents can also communicate and interface with each other according to the progress of the mandate. Parallel tasks can be synchronized, and the chain of dependence proceeds sequentially.
It was stated that multi-Agent collaboration could, on the one hand, involve different professional perspectives in the diagnosis of complex issues and enhance the quality of decision-making in mutual discussions and validation; on the other hand, it could also facilitate the continued advancement of long-link processes that would otherwise rely on manual coordination and enhance their automation and autonomous implementation.
Delegation of authority and final decision-making are retained at key judgement and decision-making stages.
Behind the three upgrades was an enterprise's own “context”
From on-site information into the system to multi-source enterprise knowledge being understood and used, to multiple Agents working together to perform complex tasks, the three upgrades link AI to the information, knowledge and implementation links required to enter real operations, allowing the enterprise context to enter more fully the task-specific judgement and execution process.
As this chain continues to operate, the knowledge, processes, rules, experience and modalities of collaboration accumulated over time by an enterprise will become increasingly integrated into the AI process, which together constitutes the enterprise's own “enterprise context”.
Internet-intelligence long-term service business communication and collaboration scenes, with an average of more than 34 billion instant communication and audio-visual services per month and an access developer of 2.35 million.
Long-term participation in the flow of information within enterprises, as well as between enterprises and their customers, suppliers and suppliers, also allows them to continue to be exposed to how business information is generated, flowing and settling, and how there is synergy between different actors.
On the enterprise AI landing path, it was argued that it was not necessary to pursue “large and complete” at the outset, and that a more realistic approach would be to move from specific scenarios and tasks to deep-seated high-value issues, gradually extending back-to-back and back-to-back, linking to a full business chain and moving AI further from single-point to operational effectiveness.
He summarized the process as a scene site, closed loop, and operational effectiveness.
Businesses can start with specific tasks such as sales, R & D, production, services, and then move up and down, creating business closed loops such as “trace to payback” “demand to deliver” “problem to solve” “strategy to implement”.
Conclusion: Business AI moves from fulfilling its mandate to participating enterprise judgement
As AI moves into more links along the business chain, the first line of information, business knowledge and business processes that it can access will become complete.
On this basis, multipleAgent can also be involved in discussions and validation from different professional perspectives, and AI ' s involvement in business analysis and judgement will be further enhanced.
In the longer term, the King Crab's goal is to further become an “engine of decision-making” for businesses.
It is argued that this does not mean that the AI substitute will make a final decision, but rather that more real business information and professional perspectives will be brought into the judgement process, helping enterprises to make faster and better-quality decisions and to retain authority and final decisions at key points.
It will be an important step forward for business AI from getting things done to helping businesses think better and judge better.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.