Manus Uses Manus to Remake Itself: Engineering Infrastructure Is Exposing Interfaces to AI

The Manus team recently shared an internal practice article about how they used AI to transform their own organization. hidecloud’s “Using Manus to Transform Manus: Practices and Reflections on AI-Native Organizational Transformation” documents a number of concrete approaches. Baoyu distilled six key points on Twitter; combined with the original article, they are summarized below.

Engineering Infrastructure: From Human-Friendly to Exposing Interfaces to AI

In the past, engineering infrastructure was centered on improving developer experience—making it more enjoyable for engineers to write code. For AI-native organizations, the first principles have changed: how can Agents be brought into contact with the real world and obtain sufficient context on their own?

Manus created a code repository called Manus-mono (Mono for short). It is not about merging all code into one place; rather, it provides a unified entry point along with a manual written for AI: what each repository does, how they work together, and what norms R&D follows. When AI needs to change a part, it pulls the corresponding repository.

The supporting capabilities include:

• Let AI run the complete Manus locally, reducing validation from a 5-to-10-minute wait to a few seconds
• Under permission controls, let AI read production logs, configurations, and exception records
• Let AI use test accounts to go through real flows such as registration and payment, recording the screen step by step
• Let AI go through the release process, while approval authority remains with humans

Whatever humans can see and operate, Agents must be able to see and operate. Only when an Agent can verify results on its own and form a “modify–verify–troubleshoot” loop does it truly move from co-pilot to driver.

The value of this step resonates more deeply with readers online. Some say that “exposing interfaces to AI” gets to the heart of organizational transformation: changing models alone is not enough; internal tools must be directly callable by Agents. Others point out that the real bottleneck for most companies’ AI-native transformation is this: who will convert documents, permissions, approvals, and other human-facing processes into versions that Agents can read?

Human–AI Interaction: Give Only Goals, Demand Evidence

In the traditional model, a human pairs with AI: AI writes a section, the human edits a section, and you still worry about whether context has overflowed and whether the code is elegant. Manus engineers do it differently: they throw the original requirement text from Slack to the Agent verbatim—tell the Agent what the boss wants, not how to do it.

If humans no longer read code line by line, how do they trust that AI’s results are reliable?

Manus uses the best models as a code review bot, letting AI prove that it did the job correctly. Product managers have the Agent use test accounts to complete registration and payment flows with step-by-step screen recording, then watch the recording and point out what is wrong.

Baoyu added a detail on Twitter: Manus product managers turn requirements into videos with AI and submit the video as the requirements document. A video can explain the scenarios in which users will use the product, how the interface is designed, what happens after a click, how different states are handled, and even the underlying implementation logic. When design, engineering, and QA watch the same demo, it is easier to align than reading documents. Baoyu said you could actually try Claude Design; for this kind of interaction, a high-fidelity prototype is more intuitive than a video.

AI Workflows: From Humans Waiting for Agents to Agents Waiting for Humans

A truly AI-native workflow is highly asynchronous and massively concurrent. With only one Agent, you have to wait for it to work; with many Agents running in parallel, you do not need to wait—you only need to switch between different Agents: either assign tasks or verify results.

One Manus engineer runs a dozen or twenty Sessions at the same time to handle requirements. Leadership throws more than twenty to-dos to a master Session, letting it do background research in Slack, Gmail, and Notion and distribute tasks on its own. A digital colleague in marketing receives tasks at just one workstation, splits them into multiple sub-Sessions running in parallel in the background, automatically gets to work when it receives a system signal, and can even raise its own request to change the system when a tool is not working well.

Extremely Flat Organization: One Person Writes All Clients

In traditional software engineering, the biggest losses occur between people: frontend and backend integration disputes, misunderstandings of documents among product/design/engineering, and misaligned cross-department schedules. Humans cannot share context at low cost, so they rely on fine-grained division of labor to compensate.

Once AI closes the technology-stack gap, division of labor is no longer an absolute necessity. At Manus, one person can write the frontend, backend, and mobile client; seven people maintain four apps; all product managers have submitted code, and half can independently ship features. One product manager took a speech-to-text feature from model selection all the way to launch—something almost impossible a year ago.

Sessions Are Becoming a New Generation of Company Assets

Dialogue streams are replacing traditional documents and Kanban boards, becoming the organization’s core work object and memory carrier. Members of the same project can see each other’s Sessions by default, so no one has to chase people for progress. Technical colleagues first complete a large amount of work with AI, then product managers enter the same Session to adjust details; two people plus AI advance in alternation.

An ops colleague got tired of always being asked to help with releases, so they wrote the release process into Agents.md. After that, anyone could have AI do the release. A small thing done to free oneself up ended up making the whole company a little faster. Readers online point out that this trick costs almost nothing and most teams could copy it today. Some also say that writing documentation for AI has a side effect: new hires get up to speed more easily. AI does not ask for context the way humans do, forcing you to say everything at once—and new colleagues benefit as a result.

When AI Does 90% of the Work, What Is Left for Humans?

Quoting hidecloud:

In an AI-native organization, the human role does not shrink; it becomes more concentrated. Define problems, judge quality, decide what is worth doing, and be accountable for results. There is also a very important point: grow together with AI.

The human position keeps moving upward—from doing things by hand to setting rules, examining evidence, and making decisions. As you move up, be careful not to hand over judgment as well. Execution can be delegated to AI; thinking should be kept for yourself.

A reader online asked a more direct question: when AI does 90% of the work, apart from the few people who define problems and make decisions, where does everyone else go? There is no answer yet.

$10,000 in Tokens per Person: Is It Worth It?

Manus’s per-person monthly AI spending exceeds $10,000, and some people use $70,000 to $80,000 in a month. What this money buys is time and attention; it buys the elimination of interpersonal communication overhead. A colleague responsible for R&D efficiency said that after the code review bot was put into use, in six months they had not heard of a single production incident caused by problematic merged code. He considers it the company’s most cost-effective expenditure.

A reader online did the math: based on Artificial Analysis’s per-question cost, Claude Sonnet 5.5 at its highest reasoning setting costs about $5.46 per question, so $10,000 can run about 1,800 questions; DeepSeek V4.1 Flash costs $0.32 per question, so it can run more than 30,000 questions—a 17x difference. Baoyu predicts that in a year or two, models like Fable may cost about the same as DeepSeek Flash. Today’s 17x gap is exactly the room for cost reduction.

Right now every team is talking about AI-native transformation, but doing it well is truly not easy. Pasting a layer of AI onto traditional processes is not transformation. Manus’s approach—rebuilding division of labor and processes around Agents in everything, and being willing to spend heavily on the best models—may not suit every company, but at least it points in a direction.