To develop an e-commerce Agent OS, "Peak Intelligence", founded by the former youngest vice president of DingTalk, has once again completed an angel round of financing worth tens of millions.

Text by | Wang Xinyi

Edited by | Deng Yongyi

36Kr learned that recently, K2 Lab has completed an angel - round financing of tens of millions of yuan. This round of financing was led by Huakong Capital and followed by Yunshi Capital. The funds will be used for model R & D, team expansion, and market expansion.

K2 Lab was established in October 2025. Just one month after its establishment, it received tens of millions of yuan in financing from Yunshi Capital. Its three co - founders all come from the Alibaba DingTalk team. CEO Wang Ming was once the youngest vice - president of DingTalk. During his nearly five - year tenure at DingTalk, he was responsible for AI innovation products, SaaS ecosystem, large - model and AI ecosystem, industrial ecosystem, and strategic terminals. Co - founder and CSO Tang Minglei had been deeply involved in the investment and research of industrial digitization and industrial AI for ten years. At DingTalk, he was responsible for strategic ecosystem and investment. Co - founder and CTO Zhao Xianlie once served as the person - in - charge of DingTalk's AI PaaS and AI operations.

K2 Lab targets the To C market and has created an Agent OS for content e - commerce scenarios called Moras. Moras serves influencers and merchants on TikTok. Users only need to interact with Moras, and it can achieve automated learning, complete product selection recommendation, script generation, content creation, intelligent editing, video pre - inspection and publishing, data analysis and other processes to achieve sales.

Currently, the user profile of Moras mainly consists of influencers and merchants with a fan base of 5,000 to 50,000. The product categories for live - streaming sales cover clothing, daily necessities, household items, festival supplies, etc.

In the past three months, Moras has completed the co - creation test with the first batch of influencer users. The data of invited test users shows that active influencers using Moras have an average monthly GMV of nearly $10,000, and some influencers have achieved a monthly GMV of over $100,000.

Specifically, some influencers have achieved a GMV of over $10,000 in the first week of registration. The order - placing rate in the first week of using Moras has reached over 70%, and this time is getting even shorter.

Moras adopts a Multi - Agent architecture and can achieve autonomous evolution. As the interaction frequency with influencers increases, the product - selection, script - writing, and analysis capabilities of this tool will gradually strengthen, and its order - generating ability will also improve accordingly.

Inspired by the Agent capabilities of OpenClaw and Claude Code, K2 Lab is now rapidly innovating and has started to establish an A2A native e - commerce operating system to serve a wider range of customers.

"Humans don't want to be in the loop."

During the invited test phase, K2 Lab found that overseas users have a much lower tolerance for the operation process than expected. Even an extra confirmation step may make users inclined to give up.

"Humans don't want to be in the loop," said Wang Ming. In Wang Ming's view, at present, AI does not have the ability to replace humans, but it can make up for human weaknesses and super - amplify human strengths.

The initial test version of the Moras product was relatively simple. However, overseas users still found it complicated, which made them realize the possibility of another business model: AI "hiring" humans.

At the beginning of its establishment, Moras designed two ways to provide services, originally intending to test which one was more suitable for the market:

The first way is that users pay a basic salary to Moras to let it play an assisting role, such as adjusting product selection and making videos. That is, humans hire AI. In this way, the platform will take a 50% commission.

The second way is that Moras takes full custody, including account assets and image. AI completely helps users create content and conduct live - streaming sales. That is, AI "hires" humans. In this model, users can only get a very low - proportion share of the revenue.

During the invited test phase, K2 Lab tested the two models simultaneously. Interestingly, more and more users chose the second business model. This means that in this scenario, as the Agent's capabilities become stronger, humans don't want to be in the loop (humans don't want to participate in decision - making).

"But the fact that humans don't want to be in the loop doesn't mean that AI can completely break away from humans," said Wang Ming. In the actual use process, after influencers and merchants authorize and log in to Moras, the platform will analyze the fan profile, the user profile of past videos, and the video tone through multi - modal understanding to push personalized product selections. Influencers and merchants still have preferences for product selection, copywriting, and videos and can review, modify, and adjust the content.

Wang Ming told "Intelligent Emergence" that Moras is conducting a gray - scale test on another version, where influencers and merchants can directly talk to the platform and tell Moras about their personalized styles, aesthetics, and ways of expression. In the future, a more complex PC - based operating system may be launched to support users in uploading professional knowledge and skills.

Frankly speaking, Moras currently has a 60 - point live - streaming sales ability, which may help influencers achieve a monthly revenue of several thousand to one or two thousand dollars, but it still has a long way to go to stably generate a monthly revenue of tens of thousands of dollars.

For an Agent, unique ways of expression such as "a sense of real - life" and "Internet sense" and sales ability are very important. To achieve these abilities, a large number of connections with humans are needed. This is also the reason why K2 Lab is self - developing a multi - modal understanding model for e - commerce scenarios.

In the content e - commerce scenario, understanding what makes a hit product and understanding users' styles and aesthetics are far more crucial than the generation itself.

Wang Ming believes that the result of AI hiring humans will be that humans find that AI can really make money, and then they will be brought into the industry, put forward more demands on Moras, and also participate in more revenue - generating processes.

To make the product closer to generating revenue, Wang Ming also revealed that Moras is training a self - developed multi - modal understanding model for e - commerce scenarios.

In his view, general models mainly optimize the generation ability and ignore the ability to understand the world. In essence, they are still "wholesaling" tokens. If a product only helps users consume tokens with a lower threshold instead of helping users improve the effect, as a wholesaler, the ROI of tokens is likely to be very low.

This also reflects the fundamental difference in K2 Lab's approach: it chooses to do things closer to commercialization - self - developing a hit - product understanding model, hit products, and hit scripts, and focusing on polishing the effect from the very beginning to let the model understand the logic of hit products.

Wang Ming said that currently, there are several content e - commerce experts from ByteDance in K2 Lab's team. After the AI learns the best practices of these experts, the AI will create the product - selection model. Human experts will become data annotators for the AI, judging which products selected by the AI are real hits. At the same time, the data results of all the content sent out will flow back to the analysis Agent and then be fed back to the previous product - selection model, forming a self - evolving closed - loop.

Users are willing to pay, the product can generate profits, and the value of tokens will increase accordingly. This forms a set of "ROI Token Economics".

Regarding the competition from large companies, Wang Ming said that currently, the top priority of large companies is still to seize the AI entrance, and the vertical scenario of AI + e - commerce has not yet entered their core vision. This is also the window period for small and medium - sized enterprises to enter the vertical scenario.

Build a Personal AI System Independently

At the beginning of this year, OpenClaw emerged suddenly. Its architecture consists of the Channel layer, the Agent layer, and the Tools layer, realizing a complete AI operating system and quickly triggering extensive discussions in the industry about Agent OS.

This has brought new inspiration to K2 Lab's products. Wang Ming said that the team also tried to develop based on OpenClaw but found that it cannot meet the stable enterprise - level use standards at this stage. Products developed with simple secondary development based on OpenClaw are still difficult to be truly implemented in the short term.

K2 Lab did not choose the route of reconstructing all products based on the OpenClaw architecture. Instead, it borrowed the design concepts of architectures such as OpenClaw, Claude Code, and Hermes Agent - including multi - layer memory architecture and dreaming mechanism - to develop a complete Agent OS independently. Specifically, they want to Agentize the Context information of super individuals and merchants in the content platform ecosystem, integrate and connect them, and form an "Agent OS" for influencers and merchants, using a solid OS system to serve the customer group well.

In Wang Ming's vision, only when the Memory of Personal AI is well - developed will the A2A (Agent to Agent) world appear. Currently, when users use centralized AI assistants such as ChatGPT, Gemini, and Doubao, they not only need to switch between multiple platforms, but the answers given by different platforms may also conflict, and the memory cannot be precipitated across platforms.

If different models in a system can be automatically called, the generated results can be automatically compared and integrated, and the historical context can be continuously accumulated, the user experience will undergo a qualitative change.

"In the future, everyone may have such a consumer - level general Personal AI assistant and a group of Agent OS for vertical scenarios," Wang Ming explained. The Personal AI system can support users to freely schedule models, add skills in natural language, and have long - term personal memory. Only when users can use the product better can they be better retained on the platform.

Based on this judgment, K2 Lab has adjusted the direction of its products - to help influencers and merchants build their own Agent OS and then build a Personal AI system independently.

In the future, K2 Lab plans to connect its products as skills to decentralized Personal AI ecosystems such as OpenClaw and Hermes Agent, which is equivalent to bringing a group of digital avatars of merchants and influencers directly to the traffic entrance of the next era.

Wang Ming believes that the continuous improvement of the Personal AI system is an important foundation for realizing A2A. In the nearer term, the next step of Personal AI is to realize a real A2A (Agent to Agent) business system.

Develop Agents with Agents

In addition to the change in the product direction, the development paradigm of K2 Lab is also undergoing a drastic reconstruction.

Although the company has only been established for a few months, it has gone through three stages of development paradigm evolution: In the first stage, AI Coding had almost no constraints, and everything was left to the AI. As a result, the speed of various bugs was too fast to fix. In the second stage, rules and structured constraints were introduced, borrowing the idea of Harness and the architecture leaked by Claude Code, using more rules to control the Agent. In the third stage, the self - developed development Agent calls Coding to develop Agents, and finally, humans conduct acceptance.

Wang Ming said that currently, more than 99% of the company's code is written by AI, and many development tasks are not completed by programmers. Product managers, human resources, finance, and operation staff are all involved in developing the system. In the latest stage, they developed a dedicated Coding Agent that can continuously help with development and iteration in a certain vertical scenario. The Agent still works on weekends when the team is resting.

Consequently, productivity is increasing exponentially. The development time for a complex requirement that originally took two weeks can be shortened to one day.

Undoubtedly, the development speed of large models still exceeds human imagination. Regarding the evolution of the Agent technology curve in the future, Wang Ming shared three judgments about the development of Agents this year: First, there will be a major breakthrough in the Memory of Agents. Second, the autonomous evolution ability of Multi - Agents will become increasingly mature. Third, Agents will code various tools by themselves to solve response problems. The era of personalized products for each individual is coming rapidly.

Regarding the future development plan, the team revealed that in the second half of the year, K2 Lab will continue to expand the team to more than 50 people and accelerate the construction of the Agent OS system. With the full - scale launch of the product, the company will also enter a period of rapid market expansion. At the same time, it will start to build an Agent - based supply chain system to take the lead in the A2A native e - commerce.