Summary

Manus’s core signal is not that it “invented Agents,” but that general-purpose Agents can now compress research, coding, and content generation into a usable end-to-end experience. Previously, users had to split tasks across OpenAI Deep Research, ChatGPT, and Cursor, rewrite prompts, and shuttle outputs between them. Now they mainly need to provide enough context and define the goal, while Manus plans and executes autonomously. Zheng Can summarizes the product philosophy as “the simplest interaction that gets me exactly the result I want.” Browser control, anti-bot capabilities, tool orchestration, and error tolerance are precisely why the “wrapper” still has value.
Models set the upper bound for an Agent’s capabilities, but whether a long-running task gets delivered depends on how well the product controls probabilistic errors. Manus’s reasoning mainly comes from its underlying models; the product’s contribution is turning an uncertain model into a relatively dependable result through planning, execution, verification, and reflection. Zheng Can uses the example of raising the success rate at each step from 85% to 90%: once multiplied across many steps, the gap in overall success rates becomes enormous. Model providers cannot take responsibility for the application, and users will not accept “GPT made a mistake, so your report is gone.”
MCP’s lead does not come from being technically optimal in absolute terms, but from making a good-enough trade-off among abstraction, expressiveness, usability, and debuggability—then organizing an ecosystem ahead of the competition. It standardizes the incompatible tool-calling formats used by OpenAI, Anthropic, and Gemini, while supporting resources, prompts, tools, and sampling, which lets tools call an LLM in return. Ya Ge’s view is that “people do not use MCP because it is technically perfect.” It remains over-abstracted in places and difficult to debug; if development stops, the de facto standard could still be replaced.
A durable Agent moat comes from the compounding interaction of tools, data, and intelligence; the most feasible and hardest-to-copy piece is the organizational ability to externalize knowledge. Tool counts can be matched relatively easily with Cursor and parallel staffing, while foundation-model intelligence is capital-intensive. If user preferences, team workflows, historical decisions, and task feedback continuously accumulate, the product develops switching costs through “understanding me better over time.” More important than the data itself is the methodology for structuring tacit know-how into best practices—an ability that is “a bit like corporate culture.”
The first Agent startup opportunities to work usually combine high labor costs, proximity to money, and outcomes that are verifiable or allow some tolerance for error. The episode highlights coding, sales pipelines, lead generation, and research: code can be run and debugged; 10 bad leads out of 100 may still be acceptable; and a flawed research report can be redone cheaply. Another clear path is converting an existing service into an Agent service—cutting delivery costs from tens of dollars or several hundred RMB to a few dollars while expanding the range of customers the business can serve.
Manus is a long-term positive for Nvidia demand, with limited near-term impact, but the real variables are the scale and efficiency of inference. Zheng Can says overseas H200 rental prices rose 10% in the month after DeepSeek launched because users could self-host and distill models. His rough estimate is that future inference usage could reach the “10x range” of training, though he explicitly calls that only a rough estimate. Today’s reasoning models think by emitting tokens and then consuming them back as context—“spit it out, then eat it.” If future models can think internally, latency might fall to one-tenth of current levels or less, but compute consumption would not necessarily decline.
Agents may first disrupt knowledge- and skill-intensive jobs rather than low-cost manual labor. People’s responsibilities will shift from personally executing how to defining what, identifying problems, and becoming “AI enablers.” Programming is already showing both productivity gains and supply-side restructuring: product managers, shop owners, and ordinary users may all be able to generate and ship software directly. The episode lands on a deeper form of scarcity in one sentence: “The ability to find problems is becoming more important”—and even the motivation to get up and work each day may become especially important.

Deep dive

1. An Agent’s minimum definition is autonomous action; the advanced requirement is self-evolution

Zheng Can considers autonomy the most fundamental property of an Agent: it can complete a task independently, or coordinate with multiple Agents, each with different strengths, rather than waiting for step-by-step instructions from a person. “The requirements for an Agent are actually much the same as the requirements for a person.”
Autonomous action requires, at minimum, the ability to perceive external inputs, call tools, and plan. Complex tasks additionally require multimodal capabilities and long-horizon reasoning. Otherwise, even if a model can answer questions, it cannot close the loop on a real-world task from start to finish.
The next level is self-evolution: an Agent should accumulate know-how and best practices from past successes and mistakes, becoming “better every time.” Koji also notes that the definition remains vague—“a thousand people have a thousand different Agents in mind”—so the episode focuses on the capability structure rather than claiming a single correct definition of the term.

2. 2024 turned Agents from a fuzzy concept into products and platforms

Ya Ge divides the evolution in 2024 into three tracks: definition, products, and technology platforms. By year-end, despite companies retaining their own interpretations, the minimum consensus around Agentic AI had broadly converged on two elements: tool use and autonomous decision-making.
On the product side, two lines emerged. Gemini, OpenAI, and Perplexity all launched research products called Deep Research; on the generation side were human-in-the-loop coding tools such as Cursor and Windsurf, the more hands-off Devin, and Gamma for generating PowerPoint presentations.
Development frameworks also expressed different worldviews of Agents: Microsoft AutoGen favors asynchronous, subscription-style message flows; LangGraph treats an Agent as something that can be described with a graph; Hugging Face’s smolagents emphasizes expressing different tasks through coding agents.

3. GAIA distinguishes between “solving problems” and “getting things done,” but does not measure long-term learning well

GAIA is a General AI Assistant benchmark proposed by researchers at Meta and Hugging Face in late 2023. It addresses a basic question: when GPT-4 scores highly on entrance or bar exams, does that reflect reasoning and generalization, or merely memorization of rules and familiar question types?
Its questions therefore emphasize reasoning, multimodality, web browsing, and tool use. The contrast Ya Ge gives is that ordinary people can reach roughly 92% accuracy, while GPT-4 with tool access scored only 15% at the time. The benchmark is trying to measure real-world task ability, not knowledge recall.
To reduce data leakage, the more than 400 questions were divided into approximately 300 test questions and more than 160 validation questions. The test set makes the questions public but withholds the answers; the validation set publishes both questions and answers for tuning. Multiple-choice formats also make automated evaluation possible.
Ya Ge points out that the SOTA results announced by OpenAI and Manus both came from the validation set. Neither submitted test-set results to the official benchmark, and neither appears on the Hugging Face Leaderboard. He sees leaderboard competition as a development or marketing tool, but GAIA mostly consists of one- or two-step tasks. It cannot capture long memory, large context windows, or “Endless Learning”—spending days or weeks understanding a codebase. Zheng Can adds that ARC is specifically designed to test genuine reasoning through unfamiliar visual patterns.

4. MCP became the de facto standard through good-enough trade-offs and an early ecosystem lead

Before MCP, connecting the same tool to OpenAI, Anthropic, and Gemini required adapting to incompatible formats. Even though OpenAI and Anthropic both use similar JSON structures, reasoning models still differ in important safety-handling details. The goal was “code once, run everywhere.”
Ya Ge says protocol competition should be judged on four dimensions: whether it abstracts away model differences, whether its expressiveness is sufficient, whether developers can use it easily, and whether errors can be debugged quickly. LangChain is his counterexample: strong on both abstraction and expressiveness, but adding a custom feature may require tracing “800 abstract classes.”
MCP describes more than tools. It also supports resources and prompts, as well as sampling, which allows a tool to call an LLM in return. Ya Ge believes it has achieved a good overall trade-off: enough abstraction without excessively damaging debuggability, backed by Anthropic’s resources, domain judgment, and stable support.
Its lead is not the end of the story. Ya Ge maintains his criticism that MCP is over-abstracted in places and difficult to debug; even with an inspector, pinpointing problems is not easy. Koji compares it with container standards: commercial and political battles began in the 1950s, and standardization only started to converge around the 1980s and continued afterward. Standards create value by reducing friction, but technical merit alone has never decided the winner.

5. Manus gets its reasoning from models; its product value comes from long-chain reliability

Zheng Can first separates two concepts that are often conflated: Manus is an Agent product, not a foundation model. Its planning, execution, and verification all depend on models, so its reasoning capability comes first from the underlying models—just as the ChatGPT product cannot be equated directly with OpenAI’s models.
The model environment changed sharply within a single quarter. Previously, o1 may have been the only model with strong reasoning; then came DeepSeek, Anthropic thinking, and Gemini thinking. The long-horizon task list Manus demonstrated benefited substantially from this leap in model reasoning.
But if every step in a 10-step task carries even a small probability of failure, the whole task can fail. Zheng Can’s example is raising the success rate at each step from 85% to 90%: a 5-percentage-point improvement in one step produces a much larger reduction in overall task failure after repeated multiplication.
Manus’s product contribution is therefore planning, execution, verification, and reflection: when it detects that it has gone off course, it replans, trying to reach the goal with as much certainty as possible on top of an uncertain model. Last year, it might plan incorrectly, reflect incorrectly, and then replan incorrectly; advances in models are what have made this loop usable.

6. Manus’s surprise was not inventing something from scratch, but eliminating cross-tool handoffs

When the episode was recorded, Manus had been public for roughly 3 weeks. Koji noted that it first went viral in China before attracting overseas attention, with Jack Dorsey posting “Excellent” and 2 articles in a single issue of The Economist mentioning it. The episode deliberately set aside the controversy for the moment: “Once the dust settles, a good product is what really counts.”
Ya Ge rejects the narrative that Manus “sprang from a crack in the rocks.” In 2024 there were already Deep Research-style research tools, along with Cursor- and Devin-style generation tools. From a feature perspective, Manus looks like a combination of the two, but the combination materially changed the user workflow.
Without Manus, a task such as “research and visualize” might first be assigned to OpenAI Deep Research, after which the user would copy the report into ChatGPT, break it down into Cursor prompts, and then generate and run the code one item at a time. The user had to design the task boundaries, divide the work among tools, and manage the interfaces.
With Manus, the user mainly explains the idea, context, and goal, then hands decomposition and execution to AI. Ya Ge also emphasizes the product’s completeness: his own data-scraping tools were often identified as bots, while Manus could still operate in environments such as Zillow and Zhihu. What truly impressed him was, “I focus on defining the problem, and AI executes autonomously.”

7. Manus’s innovation is packaging an uncontrollable model into a deliverable product

Responding to the criticism that Manus has no innovation, Zheng Can says that integrating a large number of capabilities and providing an experience that did not previously exist is itself product innovation. His definition of product power is “the simplest interaction that gets me exactly the result I want.”
This integration goes far beyond connecting a single model. A task may require a browser, image understanding and generation, data scraping, anti-bot measures, and even reading CAPTCHAs. The challenge is coordinating these capabilities while still delivering the task when any link in the chain may fail.
Zheng Can’s conclusion is straightforward: the underlying model is inherently uncontrollable, but users will not accept “OpenAI or GPT made a mistake, so your report is gone.” Cursor, Lovable, and Bolt.new are doing the same kind of work on Day One: converting a probabilistic model into a trustworthy product.
Both of Koji’s aha moments came from Manus behaving “like a colleague.” It continuously displays its reasoning, actions, and To Do List, crossing off items as it completes them. During a 10-to-30-minute run, users can ask for an update; Manus turns around to explain what it is doing and roughly how much longer it will take, without stopping the original task.

8. The compounding of tools, data, and intelligence is multiplicative, not additive

Tool compounding comes from combination. If an Agent can only search and write code, adding a report-generation tool may merely package the output. But if it can already create visualizations, presentations, and websites, adding reporting or image search may connect the entire workflow from research to publication.
Data compounding does not mean pretraining tokens; it means preferences, workflows, historical solutions, and feedback accumulated through long-term collaboration. Ya Ge compares it with an experienced technician who can identify a machine’s problem by tapping it once: a novice checks every component individually, while the veteran uses externalized experience to narrow the search quickly.
Intelligence compounds like a catalyst. A stronger model does not merely search mechanically and assemble reports; it adjusts keywords, anticipates what information will be needed next, proactively adds analytical dimensions, and may even understand that the user is asking about A while actually trying to solve B. Ya Ge says he has already felt this capability in o1 Pro.
The 3 elements form a loop: more tools generate more data to accumulate; data supports higher-level training; and intelligence improves the efficiency of tool combinations. Zheng Can uses Deep Research’s follow-up questions as an example: when it asks about report dimensions and analytical methods, it is soliciting guidance from the user and gradually accumulating “the best way to do this kind of task.”

9. Manus is a combinatorial innovation that connected years of building blocks during a window of opportunity

Koji begins with a disclosure: he is an adviser to Butterfly Effect, the company behind Monica and Manus, and an early angel investor in the team’s previous company. He has maintained a professional relationship and friendship with Xiao Hong for the past 9 years, so he acknowledges that his judgment inevitably carries personal bias.
He was not surprised at all that the product appeared; what surprised him was the scale of its distribution. The team had tried multiple Agent approaches as early as September–October 2023 and treated a Markdown To Do List as a best practice for complex tasks. By November of that year, Monica already supported web search.
In March 2024, the team built a GPTs platform. That year it also tried to build an Arc-like browser. Although the browser was never released because the team could not find a compelling commercial angle, it accumulated a foundation in browser and virtual-machine operations. Ya Ge adds that Monica already had a large collection of tuned, production-grade tools—an unusual first-mover asset.
The Roast social-distribution project in July 2024 added growth experience: it took very little time from launch to surpassing 1M UV. Zheng Can summarizes the team’s approach as “don’t play in the boundary; play with the boundary.” Koji compares it with Steve Jobs’s “connect the dots”: each building block looked thin on its own, but once the window opened, the combination catalyzed into “the shell has value in its own right.”

10. Less Structure only works when supported by stronger models and better products

Ya Ge gives “Less Structure, More Intelligence” a two-way interpretation. Reducing prebuilt scripts, flowcharts, and state machines can indeed unleash model capabilities. Conversely, weaker models require more constraints; only once models become smarter can less structure avoid sending tasks off course.
RAG represents the highly structured path: retrieve first, construct a prompt next, and have the LLM generate last, with steps and templates fixed in advance. This was more controllable when early models could not call tools proactively, but it sacrificed some creativity in handling open-ended problems.
Agentic AI allows the model to decide whether the next step is changing keywords, answering directly, or admitting that it cannot find the information. But less structure is not a shortcut. It depends on tool-use training, memory handling, context-window management, and sufficiently clear tool descriptions to reach the sweet spot of an independent thinker.
Zheng Can turns this philosophy into a startup judgment: focus more on work that leverages model capabilities, rather than piling on structures designed to constrain models. If a stronger model makes the product stronger, each new model release should be exciting. If a model improvement breaks the code originally written to constrain it, that is exactly the failure mode this methodology is meant to avoid.

11. Big tech will all build Agents, but will not all replicate Manus’s product boundaries

Asked who would launch a similar product first, Zheng Can immediately thought of ByteDance, given its existing accumulation in semi-professional Agent tools such as Coze and Trae. But he explicitly retained uncertainty: he seemed to have seen ByteDance launch a product when he woke up, yet “hadn’t verified it,” nor experienced its specific form.
Google, OpenAI, Perplexity, xAI, and others already have Deep Research products, which are themselves research Agents. Zheng Can believes all of them will probably continue building in this direction, but model companies may not want to take on detailed tasks such as report beautification and image selection. Those are more likely to be the target of application companies.
Koji also sees open-source frameworks such as OpenManus and OWL appearing rapidly. The competition will be lively, but general-purpose models, research products, open-source frameworks, and end-to-end applications have different objectives. Functional overlap does not make them the same business.

12. Coding, sales, and research will take off first because their value is high and their outputs are verifiable

Zheng Can observes that startup teams are clustering around coding, sales pipelines, lead generation, and related areas. The common thread is not low-wage labor, but large volumes of repetitive work that previously required people with know-how. Labor costs are high, the output is valuable, and sales leads are directly “close to the money.”
Coding has the advantage of being easy to verify: code can run, be tested, and be debugged. Sales leads allow some tolerance for error: if 10 out of 100 generated leads are poor, the user may still accept the result because substantial costs have been saved overall.
Research reports likewise do not need to succeed perfectly every time. If 1 report in a series is poor, the user can send it back to be redone. Given current model uncertainty, tasks that must succeed on the first attempt may not yet be mature; tasks that can be verified or cheaply reworked are better entry points.

13. Tool count is not a moat; user data and organizational methods may create lock-in

From a competitor’s perspective, Ya Ge breaks down Manus this way: with Cursor, developing an individual tool is not difficult. If a competitor invests 10 times as many people and the tools can be built in parallel, a lead based purely on tool count and breadth will quickly be matched. Tool compounding alone is not enough to defend the lead.
Effective data accumulation creates stronger switching costs. If a user corrects Manus once by saying, “Use blue consistently for company visualizations,” Manus can follow that preference automatically next time. After a long period of adaptation, switching products makes users immediately think, “Why is this new tool so stupid?” because every preference has to be taught again.
An intelligence moat requires fine-tuning an LLM, large-scale inference, capital, and systems optimization, which generally favors well-resourced companies. Small teams can still fine-tune for specific domains, however, because even a modest intelligence gain can act as a multiplier that feeds back into data and tools.
Ya Ge ultimately pushes the moat beyond data. Data can be copied; the genuinely difficult part is externalizing tacit knowledge, structuring it, and managing it efficiently. It is “a bit like corporate culture”: even if a competitor copies the tools, takes the data, and poaches the people, it will still struggle to reproduce the system-level organizational capability quickly.

14. The best startup entry point is externalizing human guidance into Agent intelligence

Zheng Can first points out that Agent opportunities do not belong only to companies building Agent products. For every startup, and even for individuals, one of the biggest opportunities of this era is to properly use already-available Agent tools and improve the efficiency of an existing business.
For those entering the field directly, the search should focus on steps that experts must guide for the work to be done well, then continuously externalize human input, judgment, and correction into the tool. The test is: after 2 years, in what dimension will the product be “unambiguously stronger” than one built by a newcomer?
Another path is to convert an existing service into an Agent service. Demand and willingness to pay have already been validated, so the entrepreneur does not first need to prove that the business exists. If a service that once cost tens of dollars or several hundred RMB can be delivered for a few dollars, and delivered faster, it may expand into customers and use cases that were previously uneconomical to serve.

15. Agents will increase compute demand over the long term, but tokenized thinking remains an efficiency bottleneck

On Manus’s impact on Nvidia, Zheng Can first gives a conditional conclusion: positive for demand over the long term, with limited near-term impact. Training may continue to grow, but Nvidia usage may not keep expanding 10x or 100x as it did in the past. More efficient model architectures are still being researched, so past expansion rates cannot simply be extrapolated.
His real-world data point is that overseas H200 rental prices rose 10% within 1 month of DeepSeek’s launch. Previously, users could only buy API access to models with capabilities similar to o1; now they can self-host and distill them. He also says the H200 happens to be well suited for deploying a “full-strength” DeepSeek.
Zheng Can believes inference will be the larger share once models are deployed, with future usage potentially reaching the 10x range of training, while explicitly acknowledging that this is only a rough estimate. Thinking models and Manus’s planning, inputs, and outputs all consume large amounts of tokens. This creates compute demand, but is also a current product-cost problem.
Current reasoning works by having the model predict its thoughts as tokens, then consume those tokens back as context—colloquially, “spit it out, then eat it.” I/O time is especially damaging to long tasks. If models can eventually perform long-horizon thinking internally, latency could fall to one-tenth of current levels or less. But Zheng Can emphasizes that this would not necessarily mean compute consumption falls in tandem; investment analysis still needs to distinguish Nvidia’s already-delivered volume from future expectations embedded in valuation.

16. The Agent era is close, with professional applications arriving from shallow to deep

Asked whether the Agent era is coming, Zheng Can’s answer is “it definitely will.” He already believed it would arrive last year; the constraint then was how soon models would become sufficiently capable. DeepSeek, various thinking models, and Manus subsequently appeared in rapid succession, bringing the era much closer within a single quarter.
Coding has become an Agent use case by default. Cursor-like products have entered programmers’ daily routines, while Bolt.new and Lovable have extended software generation to people without programming backgrounds. On the research side, Manus has appeared alongside multiple Deep Research products.
The next phase will not simply copy a single general-purpose Agent. Products will emerge across professional fields “from shallow to deep.” Zheng Can borrows the phrase “overnight, a thousand trees and ten thousand trees bloom with pear blossoms” to describe the spread: once foundation models clear the capability threshold, applications will erupt in a concentrated wave.

17. Chatbots were an accidental bonus; Agents are the main path to productivity

Zheng Can calls chatbots AI’s “accidental bonus”: something willing to chat like a person naturally has viral appeal. But if the question is how AI improves productivity, the real target has always been Agents—“actually helping me get things done, getting something genuinely useful done.”
Early products remained in chat not because chat was the final form, but because models could not do things, or could not be trusted to do them even when they could. The shift from chatbot to Agent fundamentally shows that model capability and product reliability have become sufficient to take on real-world action.
Manus’s progress stream, To Do List, and in-task conversation also mean that the interface is no longer just a question-and-answer box. Users delegate a goal, then supervise, correct, and ask questions during a long-running process instead of personally operating every step.

18. Models determine the product ceiling; productization closes the gap between capability and experience

Zheng Can believes that people ultimately use products. The illusion that “the model is the product” comes from models heavily determining an application’s upper bound, but a raw model still needs alignment, an interface, controls, and error tolerance before it becomes something ordinary users will keep using.
He uses DeepSeek-R1-Zero as an example. The reinforcement-learning model without alignment may be stronger, but it mixes languages and behaves like a “very smart stutterer.” The chat-oriented version, after bootstrap and alignment, expresses itself more fluently and controllably, although those constraints may also reduce or inhibit some of its original capabilities.
Manus probably could not have delivered today’s experience last year because the model ceiling was not high enough. But identical models do not produce identical products. Cursor and Cursor-like products, or Lovable and Bolt.new, still differ in user experience. Just as everyone can use the same batteries and electric drivetrains yet build completely different cars, the gap between capability and experience is productization.

19. Knowledge-intensive industries will benefit first and may also see supply and demand restructured first

Zheng Can continues to screen early use cases for high cost and proximity to money. Military and defense are extreme examples: as long as a technology can help, it will be deployed after various restrictions make it mature enough, even if it starts out imperfect. In civilian markets, coding combines high labor costs, knowledge intensity, and clear verification.
For programmers, Cursor is “an aphrodisiac for middle-aged and older programmers”: people whose typing speed has declined but whose judgment remains strong can become significantly more productive. Lovable also lets programmers lacking design or front-end skills produce decent interfaces. For now, coding is primarily showing up as a beneficiary.
But once those benefits accumulate to a certain point, they become a supply-and-demand disruption. A product manager may first use tools to build a demo, then bypass programmers and ship directly. A coffee-shop owner may generate an online ordering app for the store, meaning software supply no longer comes only from professional developers.
As generation costs continue to fall, software may not even need a clear commercial value and can become disposable after a single use. Writing a temporary program for a Lunar New Year red-envelope event once seemed absurd because of the cost; in the future it may feel completely natural. What changes is not merely the productivity of individual jobs, but which needs are worth turning into software.

20. Individuals will shift from rowing themselves to defining goals and enabling AI

Ya Ge summarizes the change in responsibilities as defining what rather than giving step-by-step instructions for how. Traditional programmers are like people learning to row faster and refining the material and shape of their oars. Agentic AI is like a steam engine: even the best rower cannot match its speed or ability to work 24×7.
The more valuable people in the future will not be the fastest rowers, but those who can feed coal and water into the steam engine and invent the propeller—in other words, “AI enablers.” Their core task is providing AI with problems, context, tools, and feedback so it can work as efficiently as possible.
Ya Ge uses the calculator to ease anxiety. Schools once worried that calculators would damage mathematical thinking, but after entering the workforce, people found that escaping longhand calculation and checking allowed them to solve more complex problems. He sees the arrival of Gen AI as similar to the arrival of cars, electricity, and calculators. Such transitions often happen naturally, and forming the right understanding and preparing in advance can make the process smoother.
Zheng Can adds what is distinctive about this technological revolution: it may first disrupt knowledge- and skill-intensive work that requires years of study and follows a textbook, rather than low-cost, labor-intensive industries. As the cost of solving problems falls, “the ability to find problems becomes more important.” Koji takes the point further: learning and OKRs were previously often assigned by others; in the future, “finding the motivation for one’s life” may become especially important.

Indian Premier League Cricket (Audio)Unlocking Cells' Secrets: Diffusion, Deconvolution, & Discovery with Siyu He of Squidiff & CORAL