Manus and Lin Junyang Return: AI Shifts from Models to Agents

Woofun AI reports that on August 12, the AI community witnessed a dual announcement of returns by Manus and Lin Junyang, marking a significant structural shift in the sector.

The operational timeline for Manus's transition was detailed in a notice sent to users on August 11 local time in the United States, stating the company would 'soon resume operating as an independent company.' As part of its separation from Meta, data generated by some users after December 29, 2025, would be deleted in compliance with regulatory requirements. Affected users were instructed to complete data backups by 7:59 AM Singapore time on August 23. From August 23 to August 24, the relevant data would be deleted, leaving affected accounts temporarily inaccessible.

Starting at 8 AM Singapore time on August 25, users could restore their previously backed up data and resume using the service. Manus clarified that this adjustment was not due to a data breach or security incident, but rather a measure taken during the company's transition back to independent operation to meet regulatory requirements in certain jurisdictions. Unaffected users did not need to take any action and could continue using the service normally.

A few hours later, another significant announcement of a comeback came from Lin Junyang, the former technical leader of Alibaba's Qwen project. He announced the establishment of an AI laboratory in Shanghai named Pragmatik(p7k) Labs, focusing on research into next-generation AI agents for both the digital and physical worlds. This move underscores a strategic pivot away from pure model development toward practical application.

The deeper driver is the industry's recognition that value lies in execution rather than just generation. Lin Junyang's new venture aims to bridge the gap between theoretical AI capabilities and tangible outcomes in real-world environments.

The deal between Manus and Meta began on December 29, 2025, when Meta announced plans to acquire Manus. Founded in 2022, the company initially developed in China before moving to Singapore, where it focused on creating AI agent products. Neither party disclosed the transaction amount at the time, but previous reports suggested the deal was worth around $2 billion, with a potential total value including employee retention arrangements reaching $2.5 billion. After the acquisition, Meta planned to apply Manus's AI agent technology to its own consumer and enterprise products to enhance its AI capabilities. The financial scale of the proposed merger highlighted the high stakes involved in acquiring proven agent technologies.

Woofun AI data shows that this deal later drew regulatory scrutiny. On April 28, the highly anticipated Manus M&A deal was finalized—the Office of China's Foreign Investment Security Review Mechanism (National Development and Reform Commission) issued a decision prohibiting foreign investment in the Manus acquisition in accordance with laws and regulations, ordering the parties to cancel the deal. Subsequently, Manus and Meta began working toward operational separation and stopped sharing data. According to prior reports, the two sides completed operational separation in May. Currently, Manus is finalizing data processing and preparing for independent operations. Manus stated that the deletion of data was necessary to comply with regulatory requirements.

Manus has provided data backup and recovery tools to help affected users save their task records. Affected users could perform backups multiple times during the designated window. If new task data was generated after the first backup, a new backup was required to ensure the latest data was preserved. Manus said it would not charge affected users during the backup period. After accounts were restored, the company would also offer re-entry incentives. For users registered using Apple ID or Facebook accounts, since Manus might not have corresponding email addresses, the company advised them to check in-app notifications.

The company behind Manus is Butterfly Effect, founded in 2022. Its founder, Xiao Hong, previously worked extensively in enterprise software development. Records show he studied software engineering at Huazhong University of Science and Technology. After graduating in 2015, he founded Wuhan Nightingale Technology, whose WeChat tools 'Yiban Assistant' and 'Weiban Assistant' served over 2 million business users. Butterfly Effect later introduced Monica, an AI assistant product that integrates multiple language model capabilities. Manus further expanded AI capabilities into the realm of task execution.

Positioned as a general-purpose AI agent platform, Manus allows users to issue natural language commands to enable the system to perform tasks such as information search, report writing, file processing, and code development. Co-founder and chief scientist Ji Yichao previously developed the iPhone browser and founded Peak Labs, focusing on information extraction and search technologies. Co-founder Zhang Tao had previously worked on product development at ByteDance and Lightyear Beyond. On March 6, 2025, Manus was officially launched.

Even before being acquired by Meta, Manus had completed multiple rounds of financing and experienced rapid growth. Earlier in 2025, the company raised $75 million in funding from Benchmark, with a post-investment valuation estimated at around $500 million. In December 2025, Manus stated that eight months after launching its product, its annual recurring revenue exceeded $100 million, and its total revenue reached $125 million. The company also revealed that its platform had processed over 147 trillion tokens and created more than 80 million virtual computers. When Meta acquired Manus, both parties aimed to further expand their AI agent businesses.

However, with the deal canceled, Manus's future equity structure and financing plans remain undisclosed. During the process of canceling the deal, the founding team explored raising around $1 billion to buyback company shares. Some early investors, including institutions such as ZhenFund and Hongyi Capital, also participated in discussions regarding this. Manus said that after resuming independent operations, it would continue to serve global users and planned to introduce new product features.

It aims to further enhance AI agent capabilities in the future, though no specific product roadmap has been released. Regarding data storage, Manus stated that after becoming independent, user data would continue to be stored in the United States and Singapore. Manus's return to independent operation means that this startup, centered around general-purpose agents, will once again have the freedom to independently adjust its products and financing strategies.

On the same day, another key technical figure who had previously led the development of large models in China, Lin Junyang, also focused his entrepreneurial efforts on agents. Lin Junyang announced his departure from Alibaba on March 3, 2026. He had long been involved in large model development and paid attention to issues such as inference capabilities, agent training, and model interaction with environments during the development of the Qwen series of models. The shift from inference models to agent systems indicates that the focus of AI competition is changing.

Future technological breakthroughs will come not only from the models themselves but also from how these models integrate with tools, environments, and real-world tasks. Lin Junyang has focused his new research on agent systems. The name Pragmatik Labs comes from 'pragmatics' in linguistics. Lin Junyang explained that he initially studied linguistics because a friend recommended pragmatics, after which he turned to computational linguistics and natural language processing. 'Pragmatik' represents returning to where things happen and also reflects the team's understanding of pragmatism—that AGI ultimately needs to solve real-world problems.

Pragmatik Labs is focusing its research on two areas: digital agents and physical agents. In the digital domain, the company aims to create general-purpose agents for knowledge work, enterprise operations, and industry processes, enabling AI to handle more complex tasks. In the physical domain, the company is exploring embodied intelligent systems that can enter real environments, adapt to changes, and carry out long-term tasks.

Pragmatik Labs states on its official website that next-generation agents need to possess capabilities in inference, tool invocation, learning from feedback, and coordinated action. Unlike traditional models that mainly generate information, agents need to operate continuously in environments, adjust strategies based on results, and achieve long-term goals. This approach aligns with Lin Junyang's earlier idea of 'training models → training agents.'

In his view, inference models address how to enable more efficient internal calculations before answering questions, while agent systems focus on how to allow models to take actions in real environments. Future AI systems will need to handle issues such as tool selection, task planning, environmental feedback, error correction, and multi-round task collaboration. Pragmatik Labs is currently working toward this goal, aiming to extend AI capabilities from digital information processing to real-world tasks.