Vol.55 Manus Has No Secrets — A 70-Page PPT Breakdown of AI Agents
Summary
庄明浩 sees Manus as the “visibility moment” for L3 Agents: L1 lets users watch machines generate text, L2 lets them watch models reason, and L3 lets them watch Agents plan, execute, verify, and deliver. The question “Will 2025 be the year of AI Agents?” may no longer need asking; the key point is that general-purpose tasks now have an intuitive product form: “All work is ultimately about making it visible to the user.”
Manus’s most important technical choice was not stacking Workflows, but “Less Structure, More Intelligence,” with the team setting the number of predefined Workflows at zero. The approach follows lessons from o1 and DeepSeek R1: teach less of the process, incentivize outcomes, then give the AI a virtual machine with browser, code, and API access; once too many scenario boundaries are hard-coded, a “general-purpose Agent” degrades into a vertical workflow.
庄明浩 uses 2 stress tests to argue that “general-purpose” is more than a marketing narrative: GAIA tests arbitrary-length, multi-tool tasks, while the Manus team claims it can reproduce roughly 76% of YC W25 Agent projects. In the most compelling GAIA example, Manus used YouTube’s K shortcut to pause a penguin video, took frame-by-frame screenshots to count the penguins, then pressed 3 to jump back and verify the result, ultimately confirming that the same frame contained at most 3 types of penguins; this demonstrates environmental understanding and long-chain execution, not a one-shot answer.
Agents may reshape the token demand curve first, which is why the team proposes AHPU—Agentic Hours Per User—as a complement to DAU, MAU, and retention. A normal Chat session runs for only 5 to 30 seconds, while complex Agent tasks can last more than 10 minutes or several hours; compute consumption per user may differ by 1,000x or 10,000x. The team estimates Deep Research at about $20 per question, SOTA at about $10, and Manus at about $2, while multi-task concurrency could amplify demand further.
Manus is still far from a mature product: basic information gathering is usable, but complex data extraction, source verification, visual delivery, and long-context tasks still fail frequently. When asked to identify Chinese teams among a16z’s Top 100 AI applications, it first took the shortcut of returning 9, then produced 16, 21, and 23 after repeated follow-ups before hitting the context limit; its GAIA L3 success rate is only about 57%, making the more accurate description still “a college student who works around the clock at high throughput.”
“Wrapper,” “three-hour reproduction,” and “big tech will inevitably build it” are all insufficient to establish commercial value; what remains to be tested is whether product, brand, distribution, network effects, and revenue can form a second narrative.庄明浩’s story about missing an early Perplexity investment is a reminder that building a prototype over a weekend is not the same as operating a product. OpenManus and OWL can quickly reproduce the core functions, but ordinary users still have to deal with code, environments, servers, APIs, and the command line. Overseas markets may ultimately settle the question through the bluntest test of all: ARR.
庄明浩 attributes Manus’s explosive spread to long-built goodwill and organic word of mouth rather than paid marketing, making the episode itself a stress test of the team’s distribution capability. He confirmed that none of the first on-site participants he named had been paid; the story spread from Jike, public WeChat accounts, and Xiaoyuzhou to technology, finance, and even party and state media. The longer-term user hurdle is learning to be a “good boss”: ask good questions, assign tasks, delegate fully, correct course along the way, and give feedback on the result.
Deep dive
1. Manus Makes L3 Visible as Actual Work
庄明浩 began with a self-correction: in the DeepSeek episode more than half a month earlier, the Agent section had been “a little empty,” looking more like an arrangement of existing information than a formed view. He was not deliberately withholding an opinion; at the time, neither he nor the industry had developed a clear understanding.
Using OpenAI’s L1-to-L5 framework, he placed ChatGPT-style Chatbots at L1, reasoning models such as o1 and DeepSeek R1 at L2, and Agents at L3, judging that the industry is now moving from L2 into L3.
User perception is the central thread of this upgrade: “At L1, you watch the machine generate text; at L2, you watch the model reason.” An Agent, however, must let users see that it is actually working. The original subtitle, “Will 2025 be the year of AI Agents?” may therefore no longer need to be asked.
2. General-Purpose, Low-Control, and Foolproof Point in the Same Direction
From AI 1.0’s single dataset, single scenario, and fixed model, to general-purpose foundation models and then general-purpose reasoning models, 庄明浩 asks the next question: if L3 is real, it should also be a general-purpose Agent—non-vertical, not confined to a specific scenario, and without preset boundaries—even if the early version is inevitably imperfect.
The technical path is also moving toward less human control. After o1, companies first experimented with CoT and PRM, which rewards each step of the process; DeepSeek ultimately moved from V3 to R1-Zero through outcome-only reinforcement learning, then adjusted it into R1 and reproduced the approach on Llama and Qwen. “Don’t teach it; just give it incentives.”
Product interaction has repeatedly gone from simple to complex and back to simple: natural language gave way to elaborate Prompt Engineering, while reasoning models later reduced the need for elaborate prompts. By the same logic, L3 may ultimately “not need Workflow” at all—and must be simple enough for ordinary users to operate.
3. Search and Coding Exposed the Hollow Core of the Old Agent Definition
AI search has faced the same questions since its inception: Does it need its own model? Is it merely a wrapper? Does the opportunity belong to traditional search or startups? Can information sources and product experience create a moat? Can verticalization work in legal, financial, and medical applications? Even after integrating DeepSeek, there is still no unified answer.
AI Coding faces the same set of questions. Much of the underlying progress comes from models such as Claude 3.5 and 3.7, so are Cursor-like products merely model wrappers? How much separation can engineering, user experience, branding, and vertical front-end and back-end scenarios create? “A thousand people see a thousand Agents” suggests that people had previously been describing the elephant’s legs, tail, and trunk without actually seeing the elephant.
4. A Virtual Machine Releases the PhD Intern from Paper and Pencil
Manus’s Tesla stock analysis case makes the product intent unusually clear: users can see the To-do list, web research, chart creation, and final webpage. “All work is ultimately about making it visible to the user.”
庄明浩 uses a product analogy: Perplexity gives AI search, Monica-like products give AI a browser, Cursor gives AI an IDE, while Manus “gives AI a virtual machine.” Planning, execution, synthesis, verification, and delivery can then all happen in an independent environment.
The team compares earlier foundation models to “interns with PhDs”: they know a great deal but have only paper and a pen. Give them a browser, access to internal databases, training, and clear instructions, and the boundaries of what they can accomplish expand step by step.
5. Zero Workflow Is Manus’s Most Important and Most Overlooked Choice
AutoGPT, Anthropic Computer Use, OpenAI Deep Research, and 智谱AutoGLM have all demonstrated task decomposition and execution. Reasoning models have reduced the difficulty of planning, while MCP and newer Agent frameworks have lowered the barriers to integrating tools, APIs, and architectures. The real question is now: why did it take so long to break out?
Before Manus launched, an investor or senior partner who had seen the test version asked how many Workflows the team had written. The answer was zero. 庄明浩 compares this with the shift from PRM to outcome-based rewards: the more Workflows are predefined, the more likely the product is to perform well only in a handful of designed scenarios—and the less general-purpose it becomes.
This is the origin of the title “Less Structure.” OpenAI’s core researchers have expressed the view that fewer constraints produce better model performance, while 张涛 summarized the idea in a closed-door discussion: “Less Structure, More Intelligence.”
6. GAIA Turns “Can It Work?” into a Long-Chain Stress Test
GAIA contains more than 400 questions. L1 typically requires no tools or only one tool and no more than 5 steps; L2 requires 5 to 10 steps combining different tools; L3 requires action sequences of arbitrary length, any number of tools, and broad world knowledge.
The most entertaining test asks the system to identify the maximum number of penguin species appearing in the same frame of a National Geographic-style penguin video. Manus first pressed K to pause, took and logged frame-by-frame screenshots, and arrived at 3. It then pressed 3 to jump back to the relevant timestamp and verify the result; on YouTube, the number keys 1 through 9 correspond to 10% through 90% of the video.
The team placed Manus on the same leaderboard as OpenAI Deep Research and SOTA solutions, but 庄明浩 also preserved the failure rate: GAIA L3 success is about 57%, meaning more than 40% of advanced tasks still cannot be completed.
7. A 76% Vertical Reproduction Rate and AHPU Open the Token Ceiling
About 70% to 80% of projects in one YC batch were already AI-related, while W25 produced more than 100 Agent projects. The Manus team scraped the website descriptions of projects spanning healthcare, legal, sales, finance, and even 4S dealerships, then asked Manus to reproduce them one by one, claiming roughly 76% coverage.
The team therefore proposed AHPU, or Agentic Hours Per User. A typical Chat invocation lasts only 5 seconds, 10 seconds, or 20 seconds, and at most about 30 seconds. A complex Agent may work for more than 10 minutes, several hours, or longer, making compute costs not 10x or 100x higher, but potentially 1,000x or 10,000x higher.
庄明浩 relayed the team’s figures: Deep Research averages about $20 per question, SOTA about $10, and Manus about $2. Some open-source reproductions still cost $3 to $4 to execute a single task. If users launch several tasks simultaneously, intervene during execution, or restart new processes, single-threaded AHPU will be amplified further by concurrency.
修寒 estimated at the beginning of the year that if Chinese third-party applications consumed 10B tokens per day, the cost would be about RMB5,000; several hundred such applications would imply an annual market of several billion yuan to nearly RMB10B. Even if compute costs fall 10x every 12 months and multimodality adds another 10x, Agents could still create 100x demand with ease. In an extreme video scenario, the estimate reaches 3B tokens per user; at 1M DAU, that would total 30 quadrillion tokens, roughly 300,000x the level at the time.
8. Abandoning the Browser and Publishing the Roadmap Early Show That Manus Has No Secrets
The team initially wanted to build an AI browser, but abandoned the idea on the same day Arc announced it would stop updating. A browser occupies the screen and the network, so any user action could interrupt the Agent; a virtual machine lets users assign a task and walk away, then release the machine when the task is complete.
The roadmap had already been discussed separately by 3 core members before launch. CEO 肖弘 described the combination of a cloud computer, browser, code, and APIs in February; chief scientist Peak open-sourced a reproduction of o1’s reinforcement-learning model in October 2024; and 张涛 discussed exceptional product value, observing the outside world, and getting 95% of people to use their first AI application.
The March 6 announcement called Manus “a small baby in swaddling clothes.” The team had sized its servers for an industry demo and badly underestimated the traffic that arrived within 17 hours, forcing it to enable invite codes. Model hallucinations, deliverable usability, and runtime speed all still had “a lot of room for improvement.”
9. Real-World Tests Show Basic Organization Is Usable, but Complex Data Pulls Still Take the Wrong Door
刘飞 asked Manus to score foundation models, generate reports, and organize podcast platforms, market shares, and content distributions. This kind of basic information gathering was already “quite presentable” and remains among the lower-barrier tasks.
君玉 asked it to collect official icons for several Apps; completion was about 70%, with some format errors amounting to nothing more than changing the file extension. When asked to extract monthly export values by trading partner from a customs website, it again chose the wrong data source. The conclusion is that complex research tasks still require explicit instructions about where to look.
李继刚’s Peking opera PDF was completed to a high standard. An HTML and visualization-based explanation of reinforcement learning also covered the main content, but the graphics were weak. “Cyber Star-Making Big Genius” asked it to build Doom; the result lacked gun imagery but could run, move, respond to keyboard input, and generate maps.
庄明浩’s own tests on Spring Festival travel data and a PPT exposed similar problems: the data source was wrong, diagrams collapsed into ugly stick figures, webpage screenshots still contained an unlogged-in QR code, and image placement and layout were poor. “It looks like it is working seriously; it just isn’t doing a very good job.”
10. Chasing the Same Task from 9 to 23 Shows That Agents Still Need a Good Boss
庄明浩 and 兰溪 independently asked Manus to analyze a16z’s newly released Top 100 AI applications and determine how many came from Chinese companies or Chinese teams. The task required identifying names and icons, checking company backgrounds, deduplicating 50 Web products and 50 Apps, and handling ambiguous cases such as Singapore-based teams. He judged it to be at GAIA L3 difficulty.
The control group performed poorly. 腾讯元宝 returned 10, of which 6 were not on the list. 豆包 returned 19, with 6 also absent from the list and several cases of mistaken identity. DeepSeek and 夸克 were similarly unreliable, exposing the hallucination problem in search-style products when dealing with a new ranking and multi-step verification.
Manus initially returned only the 9 companies directly mentioned in the article, clearly taking a shortcut. 庄明浩 repeatedly told it to “check them one by one” and “keep looking—there must be more.” The answer became 16, then 21, then 23, before the system hit its context-length limit. He believes the true number is higher, but 23 is already close.
This was not magic from a single Prompt. Long tasks run into exhausted context, repeated virtual-machine restarts, and pauses under heavy load. That is why the self-media blogger 伊泽’s assessment was retained: “At the level of an intern, lacking practical experience and a certain spark,” but still a college student capable of working 24 hours a day at high throughput.
11. “It Can’t Do Anything Well” and “Three-Hour Reproduction” Miss the Scale of a General-Purpose Product
Asked whether Manus is “useful,” 庄明浩 first added a qualification: almost every test has a more professional and better vertical tool, but Manus is emphasizing general-purpose capability. Borrowing Bill Gates’s old story about the internet, he asks whether the internet loses its meaning simply because radios can play baseball and car magazines can explain automobiles.
His response to “isn’t it just a wrapper?” was direct: “‘It’s just’ is a form of laziness; ‘wrapper’ is arrogance.” Manus investor 羽森 recalled that an expert once reproduced a Perplexity prototype over a weekend, and later abandoned an investment in Perplexity at a valuation of a few hundred million dollars because of that experience. A Lightspeed Venture Capital partner, meanwhile, argued that the real moat lies in product, network effects, sales channels, and brand—not in any isolated technology.
The fast reproductions by OpenManus and OWL are “very strong,” but ordinary users still have to download from GitHub, configure a local environment, rent a cloud server, connect model APIs, and sometimes run commands in the terminal. Reproducing the core functionality is not the same as delivering a stable product that everyone can use.
戈菲 uses “4-2-4” to break down the supposed three-hour reproduction: the first 4 is requirements discovery that someone else has already completed; the 2 is merely the core function; and the final 4 is long-term publicity, promotion, and operations. “Hand-building a Demo is fast. Operating a product well is hard.”
12. The Real Debate Is Whether Agents Belong to the Model Narrative or the Product Narrative
The criticism that there is “no technical innovation” contains an implicit premise: only progress at the foundation-model layer counts as innovation, so applications should not be built until model capabilities converge. 庄明浩 asks in return: if the industry also expects AI products, applications, and Agents to emerge in 2025, then what exactly is a “product”?
A former Manus employee distinguished the parts he had worked on: launching an Agent domestically for the first time between September and October 2023 and developing the To-do list; building a GPTs platform in March 2024; accumulating browser technology beginning in early 2024; and accumulating search technology beginning in November 2023. He was not involved in the social-traffic growth experience obtained through ROST in July 2024 or in the understanding of different models’ Coding capabilities developed through the Coding product in November 2024. “Each piece is relatively thin,” but combining them during the window of opportunity was still powerful enough.
OpenAI describes Deep Research as a model fine-tuned from o3, returning the question to “the model is the product.” 庄明浩 therefore asks whether the AI industry can produce a dual narrative of “technology plus product,” or whether every product capability built on top of an LLM will eventually be erased by model upgrades. “Can any innovation in products and business models ever compete with innovation in intelligence?”
OpenAI, Perplexity, and Google will all build Deep Research. According to the program’s account at the time, GPT-5 would combine o3 and GPT-4.5 while covering Pro, Plus, and Free, and OpenAI had also said it would launch a “PhD model” priced at $20,000 per month. Big-tech follow-on is “the right platitude,” and overseas markets will ultimately use ARR to decide. 张涛, however, believes competitors entering the market is itself proof that the direction is valid.
13. Explosive Distribution Came from Goodwill, Not Budget; the Final Hurdle Returns to the User
The initial spread was extremely narrow. At noon on March 5, 张涛 told people that the launch would happen that evening and the discussion the following morning. 肖弘 and 张涛 then posted videos on Jike; “Cyber Star-Making Big Genius” reposted them and tested Manus on the morning of March 6; 卡兹克 stayed up through the night and posted his experience on March 7. The first public content consisted of only 2 Jike posts and 2 public WeChat accounts.
As a participant in the closed-door discussion, 庄明浩 stated explicitly that “none of the friends at the scene took money from Manus.” The team’s announcement also said it had never opened a paid channel for obtaining invite codes and had spent nothing on marketing. The attention spread from Jike, public WeChat accounts, and Xiaoyuzhou to technology media, financial media, and eventually party and state media, exceeding the team’s expectations and control.
He attributes the organic spread to “good karma.” 张涛 had spent years producing technical explainers on DeepSeek and helping clarify rumors, making the industry easier for non-technical professionals to understand. The first friends involved were not technical users from Hacker News or V2EX, but people from their own networks. He also asks whether value really has to be validated through “exporting and re-importing” only after overseas discussion begins.
庄明浩’s personal methodology for the Agent era is to learn how to be a good boss: “Ask good questions, assign good tasks, give sufficient trust and authorization, adjust promptly along the way, and provide feedback after receiving the result—then repeat.” In what he calls AI’s “dark forest era,” if someone takes a step and takes it in the right direction, the market often delivers feedback far beyond expectations.