E201｜From Manus to ChatGPT Agent: How Do Their Underlying Technical Architectures Differ? (Part 1)
Summary
ChatGPT Agent links Deep Research, Operator, and a virtual machine into an end-to-end workflow, but its first-round experience feels more like an engineering effort to seize the general-purpose Agent entry point than a mature technological leap. 朱哲清 tested “research first, then make a slide deck,” which took 35 minutes to 1 hour; travel tasks also took 20–30 minutes. He believes the main cause is the original speed of Deep Research and Operator, which becomes even slower when the two are combined. Its advantage is using strong vision capabilities to operate on webpages, but the overall results fell short of expectations and were far slower than imagined.
For consumer Agents, the hardest problems are not clicking through webpages but personalization, payment trust, and whether the task is worth delegating in the first place. ChatGPT Agent’s selected Singapore flights and hotels were ultimately rejected because it failed to remember the user preferred Hyatt, direct flights, and lower prices; when it reached payment, a human still had to take over. 泓君 noted that human assistants also require communication. 朱哲清 acknowledged the point, but said complex preference memory remains incomplete and cold start has not been solved.
The four architectural paths for general-purpose Agents are fundamentally trade-offs among universality, speed, reliability, and permission coverage. The browser route “is indeed universal,” but comes with high token usage and network latency; an open virtual machine suits coding and data analysis but struggles with login and authorization; restricted environments trade templates for speed; Pokee-style third-party API/SDK integrations are faster and more reliable, but can execute only actions formally exposed by platforms.
Product differentiation has shifted from “who is more general-purpose” to “who can be reused reliably within a target workflow.” Manus wins on coverage through its browser-plus-virtual-machine architecture, but may be dragged down by long contexts, hallucinations, and hour-scale runtimes; JSBox breaks the super-agent into templates for slides, spreadsheets, AI calls, and other tasks, improving speed while moving toward a vertical platform; 朱哲清 says Pokee can be 4 to 10 times faster than comparable products and cut per-run tool-call costs by 50% to 60%, at the cost of giving up some personal-account scenarios.
The clearer commercial market is recurring workflows for professional users, not one-off “universal assistants” for ordinary consumers. 朱哲清 observes that many consumer Agents have poor retention because a task “ends after you use it once”; Pokee users, by contrast, repeat the same workflow every week. Standardized business travel suits API-based Agents, while leisure travel with many variables is better suited to browser-based Agents—a divide that directly determines retention and unit economics.
If Agents become the entry point over the next 1 to 2 years, portal traffic, protocol control, and content monetization could all be rewritten. 朱哲清 expects portal traffic for e-commerce, search, and video to fall rapidly. He cites Google’s launch of A to A and argues that ChatGPT, Cloud, and Pokee launching protocols likewise reflects a contest for the Agent entry point. He also estimates that of nearly 20,000 MCPs, fewer than 200 are actually usable, making maintenance and reliability hard constraints on the protocol ecosystem.
The creator economy may not disappear, but advertising and recommendation could migrate from “page ranking” to “conversational sequencing.” Responding to 泓君’s concern that AI summaries could erode podcast advertising, 朱哲清 imagines Agents paying content providers per invocation, then inserting commercial recommendations at points where there is no single correct answer. The traditional ranking mechanism that displays 5 to 10 pieces of content at once could be replaced by 5 to 10 rounds of recommendations, with each round competing for the next interaction. He explicitly preserves the uncertainty: “I’m not 100% sure this is definitely where the future is headed.”
Deep dive
1. ChatGPT Agent Leads in Browser Execution, but Minute-Scale Tasks Still Struggle to Prove Their Utility
朱哲清’s judgment after trying ChatGPT Agent on the afternoon of its launch was cool: the overall results fell short of expectations, and the system was far slower than imagined. “Research deeply first, then create a slide deck” typically took 35 minutes to 1 hour. He believes the main reason is the original speed of Deep Research and Operator, with the two layers adding further latency when combined.
A 3-day Singapore trip showed where it stands out: plan first, then search for round-trip flights and hotels, using a vision language model to understand webpages and receive feedback. Compared with systems that only scrape HTML components, this robot-like visual navigation handles real webpages better. 朱哲清 believes its clicking ability exceeds that of most browser Agents.
But the task took 20–30 minutes and ultimately stopped before payment. 朱哲清 said bluntly: “I’d rather spend some time myself and get everything done in a dozen minutes.” Payment was not technically blocked; it was “blocked at the product level”—users still do not trust Agents with their money, which is why the related capability already integrated into Pokee has not been opened.
2. Consumer Agents First Fail at Preference Memory, Not a Lack of More Operational Capabilities
The travel results exposed the difference between executing correctly and deciding correctly: neither the recommended hotel nor the flights were used. 朱哲清 usually stays at Hyatt and prefers the lower-priced option among direct flights, but the system asked only 3 or 4 high-level questions, lacked a feedback loop afterward, and ultimately delivered “a broadly defined solution.”
泓君’s rebuttal is worth preserving: human assistants also first need to know hotel, timing, price, and cabin preferences, while an Agent should theoretically be reusable after being “told once.” 朱哲清 acknowledged that communication is unavoidable. The problem is that ChatGPT remembers writing preferences relatively well, while the complex feedback from earlier travel conversations did not carry over into this task—the cold-start problem remains unresolved.
E-commerce may not be the highest-value breakthrough either. 朱哲清 believes shopping usually takes only 3 or 4 steps from a clear need to checkout, which a person can handle with a few clicks. The time-consuming part is forming preferences during the selection process, and current Agents are both slow and not necessarily capable of making choices users will accept. They do not save substantial work time in the way research, slide decks, or spreadsheets can.
3. ChatGPT Agent Is an Engineering Convergence of Retrieval, Execution, and Virtual Machines
朱哲清 characterizes the new product as Deep Research handling “where to get more complete information,” while Operator handles “how to execute based on that information.” Previously, the former delivered a massive report and the latter assumed users already knew what they wanted. The combined value is turning information acquisition through webpage execution into an end-to-end experience.
Its routing logic is not mysterious in his view: complex information requests first enter Deep Research, then carry the results into execution; simple requests go straight to execution. The virtual machine adds a third path, using Python or Bash scripts to generate PPTX files and process tasks. The overall system is therefore “a combination of deep research, the browser, and the virtual machine.”
That is also why 朱哲清 remains skeptical about the timing of the launch. The integration itself was “not actually that difficult,” making the move look more like OpenAI responding to general-purpose Agents such as Manus and Genspark and competing for the market. Neither the launch demo nor the product state appeared mature; some tasks were even slower and produced worse results than existing products.
4. None of the Four Architectures Wins Across the Board; Each Fails Differently
The first category, pure-browser Agents, assumes that every internet service ultimately appears in a webpage, so an Agent that can “see and operate” is nearly universal, with the process also visible to users. The cost is that every step may require rereading HTML and JavaScript, driving up token consumption. A single webpage download can itself take 3 or 4 seconds, making network calls an inescapable bottleneck.
The second category combines a browser with an open virtual machine. It can install packages temporarily and run Python or Bash, making it especially suitable for spreadsheet and other data-analysis tasks. It is flexible in offline computation but often cannot access services requiring full identity verification, such as a personal Facebook account.
The third uses a language model to write code and run it in a restricted environment. In JSBox’s case, the environment may come with only a small number of preinstalled packages and may not allow image-processing tools to be downloaded on demand. Its capability boundary is narrower, but standard templates for slides and spreadsheets can reduce token usage and waiting time.
The fourth follows the Zapier, n8n, and UiPath integration model, with third-party APIs or SDKs handling each node. Delivery is more reliable because permissions come directly from the platform. The limitation is equally clear: if Facebook or Instagram allows automated posting only for creator or business accounts, an Agent cannot bypass the rule for a personal account.
5. OpenAI, Manus, JSBox, and Pokee Are Betting on Different Product Boundaries
In 朱哲清’s comparison, OpenAI should still have the strongest browser-operation capabilities, especially after combining Deep Research with browser execution. In the latest browsing camp, for example, it reached scores in the 50% range on benchmark tests, while the next-best systems were only in the 20s—and did so in an open-source environment. The problem is that it “wants to put everything into the browser,” and the added capability drags down speed.
Manus combines a planning model, an independent Browser Agent, and a virtual machine to create an almost universal environment: the browser gathers information, then the consolidated output moves into the virtual machine for execution. Its theoretical coverage is broad, but overly long contexts can trigger hallucinations, and it still struggles with fine-grained actions such as uploading files and adjusting formats on complex pages. 泓君 believes 30-plus minutes is already an improvement from the 1 or 2 hours seen earlier. 朱哲清 responded that although the infrastructure is in place, browser and network-call bottlenecks remain; downloading a webpage alone can take 3 or 4 seconds.
Perplexity’s browser product takes a different route: rather than letting an Agent autonomously navigate the entire browser, it provides a sidebar while the user browses. The user specifies what should be done on the current page, and the system executes it.
JSBox is gradually breaking its “super-agent” into templates for slides, AI calls, spreadsheets, browsers, and other tasks, using fixed tools and standardized workflows to improve the experience and speed. 朱哲清 guesses it wants to refine application scenarios one by one and become a large platform hosting many small tasks; 泓君 likens it to WeChat mini programs. Because its browser navigation and virtual machine are both constrained, it is also somewhat faster than Manus and ChatGPT.
Pokee deliberately gives up some universality and relies on third-party SDKs, tool integrations, and context engineering to accelerate execution. 朱哲清 expects that after Pokee adds its own Deep Research, overall speed should be 4 to 10 times faster than market products, while tool-call costs could fall by 50% to 60%. Its overall cost could differ by an order of magnitude from OpenAI’s ChatGPT Agent and the “Manner” referred to in the original remarks, and by several multiples from products such as Verticalize JSBox. These are all estimates from his team; tasks involving unopened permissions, including personal social accounts, are explicitly out of scope.
6. Repetitive Workflows Separate Professional Agents from Consumer Agents
朱哲清 does not divide the market simply into To B and To C. Instead, he focuses on “professional users and above.” Ordinary users are unlikely to need a high-frequency, time-saving Agent, and many general-purpose Agents have poor retention because their workflows are not repetitive. The effective signal Pokee sees is users repeatedly running the same process every week.
The contrast between business travel and tourism makes the boundary clearest. Someone who flies to the Bay Area every 2 weeks and stays at the same hotel is suited to standardized automation. An ordinary tourist may change destinations, explore different hotels, or suddenly switch to business class after receiving a bonus. With too many variables, that user is better served by a browser Agent that explores and confirms along the way.
The degree of platform openness sets the ceiling for this route. U.S. technology companies generally place greater emphasis on developer ecosystems, while domestic interfaces are relatively scarce. But professional scenarios such as enterprise WeChat and creator-level WeChat accounts already support automated replies. Using Amap as an example, 朱哲清 argues that the MCP wave is forcing more companies to open their APIs and SDKs.
7. The Battle for the Agent Entry Point Turns Portal Traffic and Protocols into the Same War
朱哲清’s future workflow is that a user could ask an Agent to pull a YouTube talk by the Replit CEO and generate a growth-strategy report without opening YouTube at all. Shopping could likewise happen inside ChatGPT, from understanding body shape and trying on formalwear to finding discounts. On that basis, he believes portal traffic for e-commerce, search, and video “will fall very rapidly” over the next 1 to 2 years.
The strategic significance of protocols is therefore not technical neatness but control of the entry point. He cites Google’s launch of A to A and imagines that if one company controlled the protocol and deployed it first in Gemini, it could become the winner in the Agent entry-point race. ChatGPT, Cloud, and Pokee launching protocols similarly reflects a contest for entry points across specialized and general-purpose scenarios.
泓君 asked why Pokee would not simply join a unified MCP ecosystem. 朱哲清’s answer was maintenance quality: there may soon be nearly 20,000 MCPs on the market, but he believes fewer than 200 are genuinely usable, with most lacking active maintenance. Pokee wants service providers to hand over only their APIs while the platform handles packaging and maintenance, in exchange for an additional traffic channel.
8. Advertising Will Not Disappear, but Creator Monetization and Recommendation Targets May Be Rewritten
泓君 raised a key counterexample from the creator’s perspective: if listeners consume only AI summaries, host-read podcast ads lose their exposure, potentially breaking the commercial loop behind free content. An Agent may also cite only a small number of precise sources per task, unlike a recommendation page that can distribute traffic across 10 or 12 podcasts or videos at once.
朱哲清’s proposal is to shift payment responsibility from creators to Agents. An Agent would pay directly each time it invokes or accesses a podcast’s intellectual property, then introduce commercial ranking at points in the workflow where there is no single correct answer—for example, recommending which Agent a user should try and charging the relevant company. “The advertising industry will always exist, but the form it takes will change.”
He even believes this could improve the creator and SaaS ecosystems. Content providers would no longer have to give content to YouTube for free in exchange for an advertising split; products or intellectual property that support APIs and third-party plugins could also earn revenue per invocation. This remains a proposed future mechanism, however. The episode offered no definitive model for pricing, attribution, or covering content that gets left out.
Recommendation systems themselves may shift from spatial rankings to time sequences. Instead of presenting rankings from No. 1 to No. 10 all at once, the system could provide the most precise result—one the user is “basically certain to click”—over 5 to 10 rounds of conversation, with each round designed to trigger the next interaction. Only when relevance is comparable would it select the higher-paying content. 朱哲清 ended by preserving the boundary: “I’m not 100% sure this is definitely where the future is headed.”