The dawn of digital colleagues revolutionizing productivity is still distant.

Written by / Meng Huiyuan

Edited by / Chen Dengxin

Typeset by / Annalee

The fervor around DeepSeek has barely subsided when a new domestic sensation, Manus, has burst onto the AI scene.

Since its official launch on the evening of March 5, news about this pioneering general-purpose AI Agent has dominated the internet: "Official website traffic surpassed 10 million within four hours," "Overnight sensation on social media platforms like Weibo and Bilibili," "A-share AI agent index surged over 6% in a single day," "Relevant concept stocks closed at daily limits in batches," "Invitation codes sold for thousands, even tens of thousands of yuan"...

Amidst the excitement, stakeholders are curious whether Manus, currently in beta and not fully launched, can emulate DeepSeek's path to fame and become a new benchmark for domestic AI agents.

However, as more Manus test samples surface online, the product has swiftly navigated through a labyrinth of reputational decline, exaggerated breakthroughs, and perceived hunger marketing.

Perception overshadows actual capability

Throughout official demonstrations and multiple practical tests, the industry has gained fresh insights into the capabilities of this world's first general-purpose AI Agent.

From the official demonstration, Manus has initiated autonomous completion of intricate tasks such as resume screening and stock analysis. The entire process, devoid of manual post-optimization, directly delivers comprehensive results, surpassing similar products from OpenAI and setting a new benchmark in the GAIA test.

The plethora of online evaluations, ranging from resume screening and report writing to PPT creation and stock analysis, depict Manus users merely needing to send a task or file to swiftly invoke various tools for code writing, execution, web browsing, application operation, etc. It can also decompose tasks per user needs, executing highly complex task planning and execution.

This is the primary distinction between Manus and DeepSeek.

If DeepSeek is the "knowledge-centric superbrain" focused on optimizing language models and excelling in knowledge reasoning, text generation, mathematical calculations, and code optimization (e.g., legal contract refinement, academic paper writing, complex semantic parsing), then Manus is the "executive worker" whose core competency lies in invoking toolchains to autonomously execute complex tasks and deliver results, such as the end-to-end automation process of "crawling financial reports → writing code → deploying websites".

The previous AI sensation, DeepSeek

In terms of C-end application performance, Manus, with its cross-domain collaboration capabilities, seems poised to be the perfect replacement, liberating ordinary individuals by thinking and acting like humans.

Industry insiders analyzed, "Manus essentially integrates the functionalities of DeepSeek R1, Cline, Cloud Studio, etc. It's a relatively mature and highly operable AI Agent, but its core capability still hinges on the collaborative invocation of underlying large models and toolchains."

This has resulted in mixed reviews for Manus: Enthusiasts are optimistic about its productivity, promising cost reduction and efficiency enhancement through automated tasks like financial report analysis and e-commerce operations; critics decry its sudden errors, soaring hidden costs, and the risk of losing control in complex processes.

Mixed reviews for Manus

Given Manus's current performance, it's more apt to consider it a "digital assistant" rather than an "all-round replacement".

In this regard, Huayuan Securities believes that the significance of Manus's breakthrough lies far beyond its current actual capabilities.

It has demonstrated to the market that in the AI application 2.0 era, task-based AI/Agents should aim to execute long task sequences and possess asynchronous autonomous execution capabilities (deployed on cloud virtual machines, uninterrupted by user workflows).

This is a stark contrast to the user experience and capability boundaries of chatbots in the 1.0 era, thus justifying its position as an industrial trend node not to be underestimated.

Manus variants: The real game-changers

While the market disagrees on Manus's actual capabilities, its existence holds greater significance for the industry.

As mentioned, Manus doesn't rely on groundbreaking breakthroughs in underlying models but rather integrates existing technologies (like large models, Agent frameworks, toolchains, etc.) through engineering methods to construct a complete task execution loop akin to "building blocks".

Victor Mustar, product lead at Hugging Face, used Manus to create a small airplane game.

The MetaGPT team from domestic startup DeepWisdom recreated OpenManus using the open-source framework in just three hours; the open-source community CAMEL-AI team recreated Manus's general agent OWL in "0 days"... The key revelation from the Manus paradigm is that the threshold for Agents (intelligent agents) is not high.

In this model, the crux of Agent development lies in efficiently integrating existing resources rather than tackling technical challenges from scratch.

In other words, it's about "making the best use of existing technologies" rather than "inventing new ones." Through engineering integration, open-source collaboration, and user experience optimization, small and medium-sized teams are fully capable of creating practical Agent products.

However, as the number of Manus test samples on the market gradually increases, the limitations exposed during use also highlight the profound technical barriers in this domain.

Relevant reports indicate that TechCrunch, a technology media outlet, tested Manus on tasks including ordering takeout, reserving restaurant seats, and buying airline tickets, but errors occurred in all instances, leading to task suspension or low task completion quality.

Zhai Sen, a fund manager at Ping An, also stated in an interview with Caixin Global that he witnessed demonstrations from sellers, self-media, and other channels initially, and the effect was indeed impressive. However, due to its testing phase, resources might be limited, and it currently takes hours to generate a task for one agent.

Increasing test samples

This implies that achieving true cross-domain collaboration necessitates breaking through the core technological bottleneck—dynamic fusion and semantic alignment of multi-domain knowledge.

This entails not only overcoming the cognitive gap caused by domain barriers but also maintaining context consistency in real-time interactions. Simultaneously, it requires considering the dynamic scheduling of task objective priorities and optimal resource allocation to ensure decision accuracy, timeliness, and interpretability in complex scenarios.

This places more systematic demands on practical Agent products like Manus in terms of knowledge graph construction, context-aware algorithms, multimodal interaction protocols, and dynamic reasoning frameworks.

The testing results of Manus have also unveiled a crucial trend for this domain in the future: Competition among AI Agents is shifting from single product functionalities to ecosystem construction capabilities.

In the short term, more "Manus variants" will inevitably emerge in vertical fields, rapidly encapsulating scenario-based Agents using open-source frameworks; in the long run, to genuinely achieve cross-domain collaboration, it's essential to surmount the two inflection points—the birth of the agent operating system and the reconstruction of the human-computer collaboration paradigm.

When Agents cease to be mere tools for executing preset processes and transform into "digital colleagues" capable of autonomously understanding cross-domain task semantics, the day will truly dawn when workers' productivity is liberated.