Manus’s best product demonstration is its own team.
On October 26, Zhang Tao (online handle HideCloud), co-founder and product partner at Manus, recently published a lengthy article on his official WeChat account “Qianyun Sisi,” titled “Transforming Manus with Manus.”
The article begins with an image that left a deep impression on him. He submitted several requirements and bug reports to the engineer, and when he checked the progress a while later, he found that the engineer’s computer had multiple Codex and Manus sessions open. He casually opened one, and the input was the exact text of the requirement he had just posted on Slack—without a single word changed.
Even more surprisingly, that engineer was simultaneously working on a dozen or twenty different requirements at the time. At Manus today, this has become the norm.
This article is packed with information, and there are several key points worth noting.
Manus has built a unified code entry point called Manus-mono, accompanied by a manual designed for AI, enabling AI to view the entire cross-repository landscape. With a single requirement, one developer can implement it across all platforms. The mobile team has consistently consisted of just seven members, simultaneously managing two products and four apps, with progress on the mobile side closely synchronized with the web version.
All product managers have submitted code to Manus, and about half of them can independently bring their requirements to production. Some product managers’ deliverables have even evolved from documentation into a video.
The marketing team now has a “digital colleague” that schedules its own tasks, proactively reminds everyone of project deadlines, and even reaches out across teams to clarify functional details.
In September, the company’s AI spending exceeded $10,000 per employee, with a few individuals racking up $70,000 to $80,000 in a single month. Zhang Tao dedicated an entire session to addressing external concerns.
The timing of the article’s publication is also quite telling. At the end of September, Manus released version 2.0 and its personal AI agent app, Cue; in early October, its parent company, Butterfly Effect, announced the completion of a new funding round exceeding $500 million, while simultaneously opening 17 positions in Beijing.
From the perspective of this node, the article reads more like a manifesto marking Manus team’s entry into the 2.0 era. While it discusses how organizations can leverage AI to evolve, what ultimately emerges are insights into the company’s product roadmap, cost structure, and hiring criteria.
Manus itself is the showcase for version 2.0.
If you compare the internal practices described in the article with the feature list of Manus 2.0, you’ll find that they almost match up one-to-one.
The digital colleagues on the marketing team find the internal website inconvenient and can submit their requests directly through a feedback interface; as soon as a request is received, it automatically triggers a development task in Manus. Behind this is Automation 2.0: when an external notification is received, a task is initiated automatically.
The product manager replaced the requirements document with a video, thanks to 2.0’s significantly enhanced video capabilities, which allow both image‑to‑video generation and code‑based animation. The product manager responsible for the speech‑to‑text feature has a very typical story: she first had Manus test more than a dozen models, selected one to integrate, then built a demo according to the existing design specifications. She asked Manus to generate images for each possible user scenario, reviewed and refined them one by one, and finally edited the demo into a video. With the solution clearly defined, she simply coded the entire feature herself, designing both the web and desktop interfaces from scratch.
The Slack bot created by a colleague runs on a cloud computer. Non‑technical colleagues who want to know how a particular feature works, where the copy is located, or whether it’s live no longer need to first find out who’s in charge— they can simply ask it directly. This corresponds to the 24/7 always‑on cloud computers available in version 2.0.
There are also some features that are currently used only internally. For example, members of the same project can, by default, view each other’s sessions. After engineers and AI have completed most of the work, product managers can join the same session to fine-tune spacing, styles, and copywriting.
Zhang Tao put it very plainly at the end of his post: these are the tasks Manus needs to nail next. “We’ll test them ourselves first, then hand them over to our users.”
The phrase “eating one’s own dogfood” has been circulating in the software industry for decades. It is generally believed to have originated from an internal email in 1988 by Microsoft executive Paul Maritz, titled “Eating our own Dogfood,” which urged employees to use their company’s products more often. Later, a more famous example is Slack: originally developed as a communication tool by the gaming company Tiny Speck while working on the game “Glitch,” it ultimately became the company’s main business when the game failed to take off.
Manus takes a different approach. They build the product first, then document how it’s used internally, turning that internal usage into the product’s best showcase. For a company that sells general-purpose agents, nothing is more compelling than demonstrating, “This is how our entire company works.”
From one agent per person,
To one team, one session
In this manifesto, the section titled “The Opportunities We See” is the most thought-provoking.
Over the past year, personal agents have become standard equipment for major companies. Anthropic has Cowork, Meta’s Muse shot to No. 1 on the U.S. App Store free chart shortly after its launch, and Tencent’s WorkBuddy has surpassed 20 million monthly visits among domestic PC‑based office AI assistants. “Giving everyone an AI to do the work for you” is fast becoming the norm.
The three opportunities listed by Zhang Tao point to another layer.
First, the Session itself becomes the collaborative workspace where multiple team members work together: one person raises an issue, the designer proposes a solution, the engineer implements it, and the product manager follows up with operations—all progressing within the same Session.
The second point is cross-device, cross-colleague collaboration.
Article 3 states that once an Agent enters the company’s system and knowledge base, it can proactively identify solutions that humans have not yet considered.
All three discuss organization—how multiple people and multiple agents can work together.
This aligns with Cue’s underlying philosophy. Cue equips each Agent with its own dedicated email address, phone number, wallet, and computer. During the launch, CEO Xiao Hong remarked that such a “thing” could perhaps be called a “person,” and went on to ask, “Shouldn’t there then be a group of many people and Bots?” Cue has translated this concept into personal life, while the digital colleagues featured in the article represent how it operates within the corporate setting.
A few details about the digital colleague are more vivid than any feature description.
Many colleagues will reach out to it at the same time, so it only engages in conversation with everyone within a single “workspace” session, while specific tasks are delegated to other sessions for parallel execution. With a large volume of Slack messages, it first determines whether each message is relevant to itself; relevant messages are then categorized into two types: “requires interrupting current tasks” and “can be handled after finishing current work.” Zhang Tao specifically noted that this set of rules was designed by the AI itself during the development process.
Once, colleagues were discussing two projects on Slack and couldn’t recall which stage one of them had reached. No one mentioned it, but it proactively chimed in: “B is still in the testing phase—should we wait for A’s results before looking at B?” After the product team launched 3D model generation and rigging, it even took the initiative to ask about the applicable conditions, clarifying that rigging currently supports seven types of objects, including fish, snakes, and quadrupeds, so they could present the information accurately to external audiences.
Zhang Tao said that at that moment, it seemed more like a colleague who had been attentively listening to the meeting.
As single-agent functionality becomes standard, the next frontier for general-purpose agents will likely be team collaboration. Manus first turned his own company into a testing ground.
US$10,000 per capita
The most controversial figure in the article is the AI spending of over $10,000 per person in September. At the beginning of the month, Zhang Tao mentioned it once on his WeChat Moments, drawing numerous criticisms that the expenditure was “far too excessive.” This time, he directly provided three reasons.
First, this money buys time and attention. When a team grows large, communication costs rise; AI saves everyone’s attention, enabling each person to maintain fuller control over their own responsibilities.
Second, the cost is justified because it uses the best models—this is the core approach to understanding the limits of AI capabilities. He cited the example of a code-review bot. A colleague responsible for R&D efficiency said that in the past six months, there haven’t been any production incidents caused by merging problematic code, and since AI-assisted programming became widespread, the company’s daily code commits have become “extremely high.” This colleague believes this may be one of the most cost-effective investments the company has made.
Article 3 is the most interesting. An increasing amount of development work is being done on Manus Studio, which is a product developed in-house by Manus. Every token burned within the platform is converted into data that reveals where the waste is occurring. As the product improves, waste decreases, costs go down, and users benefit as well.
This logic mischaracterizes an expense as R&D spending. Zhang Tao’s assessment is that the work style currently costing $10,000 per month may require only a few hundred dollars in two years.
From another perspective, this calculation also highlights the shared predicament of application-layer agent companies. At the beginning of the article, the engineer had not only Manus open on their screen but also OpenAI’s Codex. Manus’s development is inseparable from the most advanced models, yet the companies that provide these models are themselves building agent products.
Some commentators have dubbed this model “dependency-driven innovation,” where one identifies a gap between capabilities and users on an established platform, at the cost of paying a fee to potential competitors for every transaction.
In this sense, the $10,000 per capita figure is simultaneously demonstrating two things: first, the team has indeed reached the limits of the model’s capabilities; and second, Manus encountered the cost challenges of Agents earlier than its users, giving it a greater chance to address these issues first.
Speed, and Zhang Tao’s unease
On the other side of the manifesto are the members of the Manus team.
In this article, nearly every role has been rewritten. Engineers have shifted from writing code by hand to orchestrating AI tasks and taking responsibility for the outcomes. Product managers are now expected to bring features all the way to production; most of the interfaces users see in Manus Studio, as well as the Video Editor, were designed by product managers themselves. Designers have also recently started writing code.
Zhang Tao summarized this shift with a phrase: moving from “Human in the loop” to “Human on the loop.” Previously, AI handled 70–80% of the work, with humans filling in the gaps; now, humans design the rules, oversee multiple AIs, make judgments at critical junctures, and intervene when necessary.
The colleague who uses AI the most on the team has the most “exaggerated” work style. He maintains a master Manus Session, solely responsible for listing the matters that require his decision and attaching recommended solutions. While everyone is enthusiastically discussing in Slack, he doesn’t even need to open Slack—he simply asks the AI, “What do I need to decide right now?”
These new definitions, in a sense, also serve as a job posting. The article concludes by stating that people remain essential, so “we’re still hiring.” Around the same time, Manus posted 17 positions in Beijing, ranging from Agent Harness engineers to business analysts; CEO Xiao Hong said they received over a thousand resumes in a single day. GitLab makes its employee handbook publicly available online, and Netflix released its culture manual early on—both approaches share the same underlying principle.
Explaining clearly how a company operates is, in itself, a way of screening for the kind of people who will join.
But what resonated most with GeekPark in this article was, paradoxically, the part where Zhang Tao expressed uncertainty.
He confessed that his feelings were very mixed—both excited and uneasy:
“Is this pace right—fast enough, or too fast?”
This question deserves further investigation. When one person is juggling twenty requirements at once, how can quality be ensured, and how can accountability be assigned when issues arise? “I haven’t heard of any production‑level outages in the past six months,” says one colleague—though this isn’t even a formal statistic. Half of the product managers are already capable of deploying features independently; how will the other half catch up? The article doesn’t elaborate, so perhaps the answers are still being sought.
There’s also a very personal passage. Zhang Tao writes that he has ADHD and over twenty to-dos every day, with the hardest part being “getting started.” Now he assigns all his tasks to a dedicated Manus project; for tasks that can be done right away, he just starts them, while uncertain ones are referred back to him. He then conducts an initial round of research using Slack, Gmail, and Notion, before handing them off to the relevant product managers.
He said this has significantly reduced his daily anxiety about “not knowing where to start.”
At the end of the article, Zhang Tao returns to that question: what do people ultimately do? His answer is to define the problem, assess its quality, decide what’s worth doing, and take responsibility for the outcome.
In AI-native organizations, the role of people has not diminished; rather, it has become more focused.
This manifesto has already demonstrated Manus’s own organizational efficiency. Whether it can become a product that others are willing to pay for depends on when and in what form those features that have so far been “for internal use only” will be delivered to users.
At the end of his article, Zhang Tao wondered what their working methods would look like if he were to revisit this topic in six months or a year.
I think that by then, looking back, we’ll probably know whether this article was a manifesto or a preview.
*Featured image source: Manus
This article is an original piece by GeekPark. For reprints, please contact Geekjun via WeChat at geekparkGO.