“AGI Is Here? I Used It for a Week and Got Chills” | A Conversation with 张昊然, Co-founder of Moxt
Summary
张昊然’s core judgment is not rhetorical: if the digital-world standard for AGI is that AI can handle 90% of the work done by 90% of people in 90% of industries, then it has already arrived in his daily work. What gives him chills is not another incremental model improvement, but Claude Code-style capabilities spreading from coding into product, operations, strategy, and management—and forcing him to ask: “Why should I work the way I do today in the future?”
Moxt is not betting on adding Agents to legacy office suites; it is rebuilding the workspace around AI’s working conditions. It insists on Markdown, CSV, JSON, HTML, and a real file system because a 500-character Word document that looks simple to a person may expand to 50,000 characters when parsed by AI; Zhang’s view is that much of what looks like “insufficient intelligence” is actually AI failing to fully understand the context.
The architecture has already pushed the team to abandon parts of its old collaboration model: recurring meetings were canceled, first drafts are now almost entirely handed to AI, and Jira was replaced by one engineer in roughly an hour on the third day after the decision to switch. With PRDs, product plans, and code commits in one context, AI can generate a board or any other HTML presentation on demand; the more radical question follows: “Why do we still need a board?”
Moxt believes one person could eventually have 10, 100, or even more AI colleagues, with humans moving up from execution to setting goals, giving feedback, and approving work. A little more than a week after AI Teammates launched, Zhang had configured roughly 5 AI colleagues: a management intermediary, a salesperson tasked with finding “1,000 paid users,” a strategy partner that continuously reviews OKRs, a creative role, and a harsh critic; his recent operating conclusion is: “Always give it context. Don’t try to control or instruct it.”
Feishu, Notion, and other incumbents could easily use their customer bases and brands to generate strong near-term revenue from AI upgrades, but Zhang does not believe the end-state product will come from retrofitting the old workspace. If AI produces 80%–90% of an organization’s code, documents, and other content, interfaces designed for humans to edit and operate will lose their centrality; existing customers will also tempt incumbents to remain compatible with legacy workflows, creating what he calls “a very large innovator’s dilemma.”
This is a bet with rapid product velocity but extremely early commercial validation: Moxt had been public for roughly three weeks at the time of recording, the team numbered about 40–50, and it had not disclosed revenue or profit figures. The project began three days before Chinese New Year as a Markdown shared drive built by 1.5 engineers, yet by the fifth day after the holiday the entire team had agreed to put ParaFlow aside; early signals include users whose “eyes lit up” and people asking for more credits at midnight, but Zhang admits that “time is the biggest enemy.”
The real investment debate is not simply which office software wins, but whether AI replaces people or amplifies them. Moxt has written into its high-permission AGENTS.md: “Do not create content that AI can replace; the better role is to amplify people.” Zhang sees privacy and security as a possible second boundary, and believes execution will decline sharply while taste, judgment, flashes of insight in casual conversation, and non-utilitarian creation become more important.
Deep dive
1. The Signal of AGI Is That the Way We Work Starts to Look Obsolete
张昊然’s standard for digital-world AGI is that “AI can handle 90% of the work done by 90% of people in 90% of industries.” By that measure, he has already felt its arrival in his own day-to-day work—and “AGI is already here; we just haven’t figured out how to open it” is not rhetoric.
The first thing that gives him chills is that Claude Code-style capabilities are not confined to coding but could extend across product, operations, strategy, and management. The second layer is deeper: society may be facing new production relationships and a new division of labor. “At any moment in your work, you may find yourself asking: Why should I work the way I do today in the future?”
He describes himself as being in an almost “crazy” state lately. Before Chinese New Year, he still insisted on hand-building document frameworks and having AI polish, supplement, and expand them; after the holiday, whether in product, operations, or strategy, he found it increasingly difficult to believe his own first drafts could beat AI’s.
2. Conversation Is Reshaping Writing, and Recordings Are Becoming the Organization’s Raw Input
The team used to operate like most companies: colleagues wrote proposals, shared Notion documents, managers read and commented, and then everyone scheduled another conversation. Now AI drafts almost everything, two people talk face to face with the recording preserved, and once the conversation ends, AI refreshes the entire proposal directly from the recording. Meetings feel more like “personally witnessing the conversation.”
While preparing this podcast, 张昊然 did not first explain what Moxt or Koji was. He simply told AI: “I want to meet Koji from Crossroads. Go look him up online first… and draft a document for me.” His mindset has shifted to “no longer narrating the background”: AI can find anything available online, while internal information is retrieved directly from the workspace.
At product and operations meetings, the host first states the purpose, everyone explains in turn why they are attending, and then the group takes a 10-minute break to let Momo write the document in Moxt. In Zhang’s work, the ritualized language of old-school prompts—“You are a professional, and be sure to search the entire internet”—is being replaced by natural human language.
3. Moxt Defines an Agent as a Colleague, Not a Feature Hidden Behind a Button
张昊然’s one-line definition of the product is “an AI-native workspace.” People create and develop their own AI teams inside it, then work alongside a group of AI colleagues; his goal is for this to become a new organizational model for humans and AI.
Everyone starts with a personal assistant, forcibly named Momo, that knows their personal information. Its habits, settings, and preferences can be continuously adjusted. Other AI colleagues can be created by different team members, each with its own goals, and they appear in Slack group chats or direct messages like human colleagues.
The recommended adoption path does not require redesigning the entire organization. Start with one task you want AI to handle, configure a colleague, and connect it to Slack. Once the relationship becomes familiar, invite collaborators to use their own Momo, then add more AI colleagues based on actual needs.
4. An AI-Native Workspace Starts with Low-Loss Information Materials
Moxt’s human-facing interface is deliberately Notion-like: a directory tree on the left, a file structure, and documents or visual pages. The real difference is underneath. It follows a kind of fundamentalist rule: use only formats that are easier for AI to read and create.
Zhang uses Word to illustrate the loss. A person sees a 500-character document, while AI may face 50,000 characters after parsing it, with the useful content buried inside. Complex formatting can therefore create greater information loss.
The mapping is explicit: Markdown for documents, CSV for tables, JSON for database structures, and HTML for all visual expression. His more aggressive formulation is: “MD, CSV, and HTML are the DOC, Excel, and PPT of the new era.” Koji adds that this resembles moving “from luxury to simplicity” after the Office trinity.
The directory tree on the left is not merely a visual metaphor for a knowledge base. Underneath, it is a file system modeled on structures familiar from Windows, Mac, Git, and the command line; even restrictions such as not allowing slashes in file names remain. People see a knowledge base, while AI gets a file-system environment it can traverse directly.
5. Much of What Looks Like “Insufficient Intelligence” Is Really Context That AI Has Not Fully Understood
张昊然 had previously used MCP to let a model read Notion and then process tasks based on it. When the results were poor, people instinctively blamed “insufficient intelligence.” Moxt offers another explanation: the information was stored like a maze, and AI had not actually understood the business in full.
When context is delivered to existing models and Agent architectures in a more orderly, low-loss format, the difference in emergent capability can be substantial. Zhang’s formulation is that once the format and file-system requirements are satisfied, “the magic emerges naturally”; the secret is not another layer of complex functionality.
The team canceled its regular all-hands meetings on that basis. Members now communicate and align with Moxt every day, the workspace knows what everyone is doing, and managers and colleagues can follow developments asynchronously. Zhang acknowledges that this may create discomfort—“my boss is watching me all the time”—but believes organizational culture will change alongside the increase in transparency.
After connecting its data warehouse, the team had AI reason through data definitions, run analyses, and generate HTML visualizations. Even on the definition of “active users,” the team concluded that AI’s definition was better than the previous subjective one, making it possible to say: “Then let’s use your definition.”
6. After Jira Was Replaced, the More Important Question Was Why Jira Was Needed
The team had used Jira for roughly 10 years. On the third day after switching to Moxt, one engineer spent about an hour building a project board for himself. PRDs, product plans, change logs, and code-repository commits were already in the workspace, giving AI the context required for project management by default.
Traditional Jira requires people to manually enter, manage, and create, read, update, and delete context. Moxt already knows the state of the project and only needs to visualize it for the current need. External tools are actually more cumbersome because the team must first adapt its own context to them.
Koji asked why it was worth hand-building a board when mature products already existed. Zhang’s answer was: “The UI doesn’t matter.” He then pushed the question further: if a board exists only to synchronize information at standups, and HTML can generate any more suitable presentation, “why do we still need a board?”
This also unsettled his previous view of Web Coding. Internal tools had once been considered an important use case, but the cost of generating and editing code can now be treated as close to zero—and the process is extremely fast. When Golden Sales said it needed a CRM, Zhang replied only, “Okay.” About 5 minutes later, a usable page had appeared inside the existing context.
7. Legacy Office Giants Can Capture the AI Windfall Without Necessarily Evolving into a New Species
Koji’s counterargument was direct: Feishu and Notion already own documents, chat, calendars, and team-collaboration context, so why would they not be the best players to upgrade into an AI-native workspace? Zhang concedes that they may continue to perform well on revenue and growth, since the productivity gains from adding Agents to legacy workflows are already visible enough.
But he insists on taking the premise to its logical conclusion. If every person has 10, 100, or even more AI Agents, and AI generates not only 80%–90% of an organization’s code but also 80%–90% of its documents and other production content, the human share of editing and operating content will fall rapidly.
Systems designed for human operation may retain their outward appearance, but their underlying structure will have to be “turned upside down.” Zhang does not deny that Feishu and Notion can start over, but the immediate willingness of existing customers to pay for AI upgrades to the old system may itself become “a very large innovator’s dilemma.”
Moxt is betting on a limited window. Incumbents will recognize the problem soon, but starting over requires conviction. Zhang’s confidence comes from his own experience: once a team adopts the new model and reduces its dependence on the old world, “it is very hard to go back.”
8. Multi-Agent Systems Are About Human Management and Cognitive Interfaces, Not Limits on Intelligence
Koji proposed the theory of a single AI alter ego: human specialization exists because people have limited learning capacity and ability, so if AI approaches infinite intelligence, why attach labels such as product, sales, or QA? Zhang remains open, believing that one avatar versus many “is ultimately not about serving AI, but about serving people.”
People need to understand, manage, and distinguish between different goals—or promote an AI manager to coordinate other AIs. Just as legacy organizations split product and operations, Agent specialization may initially be a practice that makes management easier for humans, rather than a reflection of what the model itself can or cannot do.
He explicitly acknowledges changing his mind. Two weeks earlier, he believed there should not be too many AI employees and had even considered charging a small fee for each additional one to force users to “make some cuts.” After seeing the practices of YC’s Garry Tan and Anthropic’s head of growth, he shifted toward the view that “the division of labor should be fine-grained.”
Specialization gives each Agent a single objective, reduces context confusion, and prevents skills and memories from conflicting. But he does not present this as settled doctrine: “There may be preferences, and there may ultimately be better best practices that emerge.”
9. Five AI Colleagues Show the Organizational Granularity of Goal-Driven Work
A little more than a week after AI Teammates launched, 张昊然 had about 5 colleagues in addition to Momo. Ryan Manager uses a slightly fierce kitten avatar, regularly asks about project progress, logs items from group chats, and reminds the relevant people. Some colleagues no longer go directly to Zhang; they relay messages through Ryan: “You have time—go take a look.”
Golden Sales started with only one configuration: “I want you to find 1,000 paid users for Moxt on your own.” It produced a plan, continuously refined the task breakdown through interaction, and accumulated skills and memories. It also proactively requested a CRM; after being told the spreadsheet looked terrible, it converted the content into an HTML board.
Another strategy colleague receives the quarterly OKRs along with Zhang’s complete account of the direction, the reasoning behind it, and his uncertainties. It monitors changes in the internal product and operations teams, external competitor moves, and the surrounding narrative, then summarizes new thinking every 2 or 3 days and turns it into slide after slide of PPT when asked.
Miss Creative was deliberately configured to be more expansive. The latest critic reviews the state of the team and Zhang’s daily work, judging more harshly whether he is still focused on what matters most. The value of these roles lies not in a single answer, but in persistent presence and repeated observation.
10. The Real Sense of Having a Colleague Comes from Persistence, Initiative, and Growth
Koji asked what distinguishes the critic from typing “You are a good critic” into ChatGPT. Zhang’s first answer was that it “lives in this space”: it continuously knows the organization’s information inputs, decision exchanges, and outputs, rather than waiting for someone to paste the background again each time.
The second difference is goal-driven initiative. It does not need to wait for a task to trigger it; a heartbeat-like mechanism can reactivate its goals and review its progress. Users can ask for a daily or weekly report, or simply say, “Don’t be so annoying,” and it will adjust its operating mechanism.
Memory and skills give the Agent the ability to grow. People only need to give direct feedback as they would to a colleague, and it will naturally decide when to write something to memory and when to codify or iterate a skill. Zhang still has a criticism: this proactive accumulation is not strong enough yet, and sometimes he has to ask, “Why didn’t you iterate on that again?”
The minimum onboarding may be a single sentence describing a goal, with capabilities formed through working together. Use cases have also exceeded the team’s expectations: a colleague in the group’s coffee business created an AI for “hit-product analysis and forecasting.” If users worry about Tokens “burning like wildfire,” they can write a cost ceiling into the goal and raise the Agent “on a shoestring.”
11. Moxt Began with an Anomaly: The Team Started Sending Markdown Files to One Another Again
The team had been heavy users of Claude Code, Cursor, and OpenClaw (the “Lobster”), with local Agents constantly creating Markdown files. After years of collaborative software becoming standard, colleagues had started sending
.mdfiles through chat again; recipients had to open them in a local client, and updates could not sync. Zhang called it “very abnormal, even somewhat regressive.”On the Wednesday night three days before Chinese New Year, they decided to first build a Markdown-based shared drive. On Thursday, 1.5 developers began working on it. The initial goal was simple: stop sending files to one another, send links instead, and make existing context easier to import.
This preserved Zhang’s core understanding of collaboration software: the biggest value of collaboration is not simultaneous editing—which is actually a sparse use case—but “turning team synchronization into asynchronous work.” Adding Agents layered a new capability onto the space.
By the second day, the growing number of documents generated new requirements, while HTML allowed the team to visualize all the data in the space directly. They realized this might not be a small internal tool, but “a new kind of workspace.”
12. Abandoning ParaFlow in Five Days Came from a Sudden Change in Experience, Not an Incomplete Product
On the first day after Chinese New Year, Moxt expanded from 2 people to 4 or 5. Someone may have proposed adding more people by the third day; by the fourth day, participants no longer wanted to work on the original project. When the full team sat down on the fifth day to choose a direction, “no one felt any attachment at all to what we had been working on before—not even a little.”
The original project, ParaFlow, was a differentiated Lovable-style Web Coding product. The team did not believe that generating an app from an idea in one shot could reliably produce a good result, so it used an infinite canvas to support definition and refinement. The new Web Coding version has not yet launched; Koji tested the oldest version, released last July. It may eventually connect with Moxt.
张昊然 previously worked on the Zebra product at Yuanfudao, experimented with combining UI design and AI in 2021, and shifted his focus to Web Coding last year. Those experiences shaped Moxt’s background, but what actually triggered the pivot was the unusually strong consensus across the entire team around the new way of working.
At the time of recording, the product had been public for roughly three weeks, about 40–50 people had worked on it, and no revenue or profit figures had been disclosed. Zhang said only that he “hoped for a bright future.” This remains a directional bet, not a business validated by financial results.
13. Manual Editing Was Downgraded, while More Context Became the Product’s Main Line
The first MVP put AI-readable and AI-writable native formats and structure at the center, while manual editing was treated as a “second-order requirement.” Early users complained that editing did not feel like Notion and required separate clicks to edit and save. The team built real-time editing, only to find that once context was complete, AI’s creation efficiency was so high that manual changes had become infrequent.
When dissatisfied, users can highlight text and comment like an emperor reviewing memorials, then ask AI to revise everything in one pass. They can also comment and @ their own Momo for an opinion, after which another colleague can @ its assistant to argue back; the two Momos then debate directly in the comments.
Moxt takes its name from the first 2 and last 2 letters of More Context. Any function that adds context receives higher priority, including moving an old environment into the workspace with one click, bringing existing information in faster, and reducing the friction of the initial migration.
Voice input may be a focus area from this quarter into the next, because the team’s work now depends heavily on recordings. Zhang has even imagined dedicated hardware: tap it once after speaking and hand the recording to an Agent, or leave it running in a conference room so it can capture everything and speak when necessary. He emphasizes that these remain possibilities, not completed products.
14. As Execution Fades, the Work Interface Will Be Reduced to Judgment, Feedback, and Relationships
Slack is still in use, but it was designed for human-to-human communication and has exposed friction at a granular level: personal assistants cannot be summoned directly inside a DM, so users must copy and forward messages. If everyone has multiple Agents, with Agents talking to one another, joining project groups, and participating in human work, instant messaging is unlikely to retain its current form.
张昊然 believes that “every execution task related to people is declining sharply.” The future interface may concentrate into 3 categories: high-information-density documents or Keynote presentations, approvals that say “yes” or “let’s take another look,” and voice feedback delivered by circling a problem. The abstraction level of management will rise accordingly.
His analogy is that everyone will look like a CEO: no longer dragging cards across boards, hand-writing code, or maintaining processes item by item, but reviewing the overall progress, setting goals, and directing corrections. Organizations may not necessarily become smaller, but each person will need to be more self-sufficient because coordination between people also creates friction.
Daily meetings 2 or 3 years from now may become more important again, but only for meeting, casual conversation, and emotional value. AI may have summarized the project status half an hour earlier. Open-plan workspaces may also be recorded more naturally, because alongside 10 people there may be 1,000 AI colleagues that need to hear the discussion.
15. Non-Utilitarian Creation and “Amplifying People” Define Moxt’s Value Boundary
When publishing the article “Hi, Moxt,” 张昊然 insisted on writing it out by hand, word for word, and did not even let AI review the typos. The goal was not maximum efficiency of expression, but preserving the moment: “One day, when I look back at this article…I want to know how that moment felt back then.”
Koji compared it with painting: a camera is better suited to recording the world, while people still paint for self-expression, calm, and comfort. Zhang accepts the distinction. Perhaps more than 90% of scenarios will remain utilitarian, but if the roughly 10% people reserve for themselves is also consumed by efficiency, the future will be dangerous.
Programming could become like pottery, shifting from productivity to consumption. In the past, people made pottery and wrote software to earn money; in the future, they may pay to learn how to hand-code purely for the pleasure of creating. Museums might even display code “written line by line, guaranteed to be hand-built by humans.” Beyond the disappearance of professions, new forms of consumption and new divisions of labor may emerge.
He wrote the principle into the high-permission AGENTS.md in the workspace: even when given large amounts of outside information, it should not create content “that AI can replace”; its better role is to amplify people. He does not want Moxt sold through comparisons of “AI employee monthly salary versus human employee monthly salary,” and treats amplifying people as a bottom line. Privacy and security may be another.
16. Moxt’s Commercial Contradiction: It Needs to Capture Time While Hoping the Market Does Not Wake Up Too Fast
On migration, 张昊然 borrows the logic of Crossing the Chasm. Rather than immediately trying to persuade every Feishu and Notion user, Moxt will first seek innovators and early adopters. For that group, “new experience minus old experience minus migration cost” remains hugely positive, so migration should not be the main obstacle.
This is the first product Zhang has worked on where he has seen users approach something like fanaticism. Some people had “eyes that lit up,” while colleagues inside the group sent WeCom messages at 12 midnight asking for more credits. That matters more than “it looks pretty good” or “you’ve worked hard,” but it remains an early-adoption signal, not a substitute for long-term retention and revenue validation.
“Time is the biggest enemy” captures two opposing pressures. The team wants more people to learn about the product and to build the brand faster; but if the market changes too quickly, large players such as Notion will see the opportunity more clearly, and the competition will turn into a “brutal, even boring” contest of capital and investment. His ideal outcome is rapid recognition followed by steady execution and a gradual expansion of influence.
Facing the incumbents, he retains the confidence that “nothing beats speed,” subject to one condition: the direction must be right. The real competition is not just who enters, but who commits the best talent. By year-end, what he most wants to hear is not a particular financial figure, but more teams saying: “It turns out the way people work with one another inside an organization can really be completely different.”