Summary

Industry buzz around Agents has yet to translate into real demand, with both supply and demand voting against “all concept, no solution.” Qunxiang has more than 10,000 members, and its previous sessions regularly drew 2,000—3,000 attendees, yet only 900 people signed up for its AI startup course on day one; a US B2B survey cited by 庄明浩 found that 94% of companies were dissatisfied with their AI vendors. “Everyone knows AI is a treasure,” but most services remain stuck at “monetizing anxiety and explaining some logic.”
A consensus path from L1 to L2 has emerged—general-purpose, lightly controlled, and rewarded only for outcomes—but whether that playbook can take the industry to L3 remains the biggest open question. OpenAI’s ladder runs from L1 chatbots and L2 reasoners to L3 Agents, L4 innovators, and L5 organizers, with leading companies now moving from L2 toward L3; 庄明浩 believes the consensus that 2025 will mark the shift from enterprise experimentation to real deployment, with Agents delivering the answers, still contains considerable wishful thinking. Reinforcement-learning work such as DeepSeek R1 supports “less control, more intelligence,” but that does not mean Agents can dispense with boundaries, workflows, or human-designed structure.
The biggest investment risk at the application layer remains that a single model upgrade can swallow years of product capability built by a startup. 归藏 argues that GPT-4o’s image generation can cover almost every complex Stable Diffusion workflow through natural language, again suggesting that investors should be cautious with applications before capabilities converge; that does not make products worthless, however, and Manus’s “show your work” paradigm—planning, to-do lists, step-by-step execution, synthesis, and complete delivery—is becoming the interaction language of L3.
Agents still lack a unified definition, but the industry is converging on four traits: memory, planning, tool use, and action. The real leap is not better conversation but “from language to behavior”: calling existing tools is already enough for most general-purpose tasks, shifting the hard problems to context, permissions, data and security architecture, and stitching search, browsers, business software, and code execution into end-to-end outcomes. The underlying distinction between Deep Search and Deep Research will ultimately matter less to users than result quality.
The opportunity structure is beginning to stratify: general-purpose Agents and platform ecosystems look tilted toward giants, while early-stage capital is concentrating relatively more on specialized Agents; infrastructure remains unfinished, which is why so many projects are emerging. Browser Use raised $17M on March 22, Browserbase topped the early-stage Enterprise Tech 30 ranking, and A1Base provides Agents with nothing but Email, WhatsApp, and phone-number identities; the fact that the value chain is “split into so many specialized layers” is itself a signal that the Infra layer remains unsettled.
“If the boundaries are unclear, call it an Agent; once the boundaries are clear, you stop calling it one”—which means conceptual valuation will ultimately give way to concrete business value. Under that deliberately provocative framing, AI search, AI Coding, and Deep Research may stop emphasizing the Agent label once their delivery boundaries are clear; a so-called WeChat Agent may have the latent advantage of Mini Programs as its “hands,” but privacy, permissions, and WeChat’s cautious rollout cadence remain constraints. “The last concept that tried to contain this much was the Metaverse.”
庄明浩 believes Agent monetization may not stay at $20 per month for long, with task-level costs, ARR, and automation value likely to face market scrutiny sooner. Manus charges roughly $1 per 100 credits: 15 minutes of data analysis costs about $2, 25 minutes of web development about $3.6, and 80 minutes of application development about $9, forcing users to decide “which tasks are worth spending a few dollars to execute”; with inference still expensive, complex environments unreliable, and users still required to supervise, automating ordinary white-collar work may prove more profitable than Nobel-level scientific reasoning.

Deep dive

1. Buzz Has Yet to Become Demand; Concepts Are Exploding Ahead of Solutions

庄明浩 opened with Qunxiang’s registration gap: it has more than 10,000 paying members, and its major courses typically draw 2,000—3,000 attendees, yet its AI startup course attracted only 900 people on the first day despite a full promotional push—well below the original expectation that “3,000 would fill it on day one.”
刘思毅’s explanation was preserved in full: “Everyone knows AI is a treasure,” but entrepreneurs also understand that, beyond anxiety marketing and explaining logic, the market still has not found a genuinely useful set of solutions for them.
The other side is no more encouraging: 庄明浩 cited a US B2B enterprise survey showing that, looking back at 2024, 94% of companies were dissatisfied with their AI vendors, leaving only 6% satisfied. At the same time, he noted that the market should by now have produced something like Manus for Beginners, creating an absurd contrast.
Zhipu has also released AutoGLM 沉思, an AI Agent resembling Deep Research; Agent, Deep Research, AutoGLM, GLM, and 沉思 piled up layer after layer, leaving non-industry users unable to understand what they are looking at and exposing a major language gap between technical, product, and market teams.

2. The L1-to-L2 Playbook May Not Transfer Directly to L3

OpenAI’s AGI ladder runs from L1 chatbots and L2 reasoners to L3 Agents, L4 innovators, and L5 organizers; 庄明浩 judges that the world’s leading companies are broadly at the point of moving from L2 to L3.
L1’s breakthrough was generality: from rules, machine learning, neural networks, and Transformers to GPT-3 and ChatGPT, models stopped relying on a single domain dataset to perform just one task. The previous generation of AI ultimately clustered around vision, cameras, and security precisely because its use cases had been narrowly defined.
L2 came from more complex inputs, chain-of-thought, and post-training. The industry once debated whether to reward every intermediate step or only the final outcome; subsequent practice pointed toward reinforcement learning that rewards outcomes without imposing structure. DeepSeek used V3 as the base for R1-Zero, then improved usability through post-training such as cold-start data, and replicated the approach with Llama and Qwen.
Two principles have therefore become consensus: “general-purpose,” and “less control, more intelligence.” 庄明浩’s real question is whether L3 should also avoid restricting scenarios and imposing workflows, relying only on outcome-based incentives. The episode did not pretend to have a definitive answer.
2025 was expected to move enterprises from AI experimentation to real deployment, with Agents taking over the work; 庄明浩 believes that convergence around this timeline still contains a good deal of wishful thinking, and that room remains for scenarios and verticals.

3. Model Upgrades Are Still Eating Application-Layer Moats

The main story over the past 2 years has remained the use of more capital, infrastructure, and training to produce stronger models. When DeepSeek exploded, the product format was not fundamentally innovative; the core was still “intelligence plus open source.” That has kept 庄明浩 questioning whether product and business-model innovation can ever consistently beat intelligence itself.
GPT-4o’s image capabilities set off a wave of Ghibli-style content. 归藏 argues that it can cover almost every variation of a complex Stable Diffusion image-generation workflow through natural language; some people were studying Stable Diffusion tutorials one day and discovered the next that “all of it had been a waste of time.”
庄明浩 cited 雨森’s Perplexity example: engineers may believe they could reproduce it in a few days, but if they had invested hundreds of millions of dollars early on, Perplexity would now be a company worth more than $10B. Whether it can kill Google is a separate question; the early investment return is already real.
But calling something a wrapper also understates the engineering difficulty. Even a thin product layer on top of a model still requires complex work from defining demand to delivering reliably; AutoGLM 沉思’s underwhelming implementation shows that knowing how something should work and building a good product are two different things.

4. Product Trends Are Starting to Emerge; “Show Your Work” Becomes the Agent’s Interaction Grammar

ChatGPT’s magic moment was watching the machine produce text; DeepSeek R1’s was watching the model think. Manus’s definition of L3 pushes the logic further: users should be able to see the plan, to-do list, web and code operations, research organization, synthesis, and the final delivery of a complete webpage or report.
When Metaso Search turned a long answer into an interactive webpage, it deliberately displayed scrolling HTML code and a progress bar moving from 0% to 100%. Generation itself is not mysterious, but the product chose to let users “see the process,” turning waiting into visible labor.
戈飞’s “4-2-4” principle captures the point: technical implementation is only the middle “2”; demand and product boundaries on the front end, and operations and promotion on the back end, each carry a heavier “4.” That is the real distance between an open-source solution and a product users can actually use.
Enterprise Tech 30 offers weak but clear corroboration: the share of AI-native companies rose from 0 in 2019—2020 and 3% in 2021 to 33% in 2023, 40% in 2024, and 50% in 2025; non-technical PLG also rose from roughly 15%—20% to 25%, suggesting that product, sales, and brand narratives are making a modest comeback.

5. “Agent” Is Still the Blind Men and the Elephant; Workflows Are Both Patch and Boundary

Games have long had scripts, machine learning, and autonomous behavior: teammates and opponents outside the player’s control in football games, and even the 4 enemies in Pac-Man with different rules, were all used by 庄明浩 to ask whether they count as Agents. Today’s debate is really about the large-language-model generation of Agents, not the first appearance of Agents.
Looking back at YC projects from 2023 through early 2024, 庄明浩 noted that Agent was already the seventh-largest category. Projects used screen recording, drag-and-drop, natural language, RPA, logic trees, and workflows as proxies for Agents, yet differed very little; “another consensus Agent” may simply be an intermediate state.
The logic at the time was “the model is not capable enough, so use workflows to make up the difference.” But workflows cannot enumerate every situation; they can only constrain the scenario and its boundaries. The industry has therefore cycled repeatedly from simple to complex, then simple again, then complex again: prompts and chain-of-thought followed natural language, and workflows were piled onto Agents after reasoning improved.
GPT-4o’s ability to cover complex image workflows with natural language again suggests that path dependence may be leading humans down a detour. If L3 continues toward generality and less control, it may ultimately also need to move beyond workflows exhaustively designed by humans.

6. What Is Converging Is Not the Definition, but Memory, Planning, Tools, and Behavior

庄明浩 compresses the current consensus into 4 features: memory, planning, tool use, and action. Memory remains constrained by the context window, and Manus’s frequent context-limit errors show that this foundational problem is still unsolved.
Planning once depended on “tactical effort” such as workflows; it only began to look viable after reasoning models emerged. On the tool side, there is no need to rush into inventing new tools: the existing ecosystem can already handle most general-purpose tasks. The real problem is invoking them reliably.
Tool use continues to create problems around access permissions, data architecture, and security architecture. So-called end-to-end training has not eliminated boundaries; it has created a new engineering system that combines necessary constraints, reinforcement learning, reasoning, and connections to external systems.
He distinguishes Deep Search from Deep Research: the former is built on search, while the latter is described as an end-to-end-trained model. Users, however, do not care whether the underlying system is RAG or end-to-end; they care about result quality.
The explanation for games’ poor real-world performance is straightforward: “This wave is built on large language models—language. But games need behavior.” From Computer Use to OpenAI Operator, leading vendors are all trying to make language models act in the real world. That is the core transition from L2 to L3.

7. General-Purpose Opportunities Go to Giants; Early Capital Favors Specialized Agents and Infrastructure Projects Abound

CB Insights’ Agent map has 3 layers: Infra; horizontal work such as software development, data analysis, security, sales, HR, and Marketing; and vertical industries such as legal, gaming, and healthcare. The segmentation also maps to different avenues for capital deployment.
General-purpose model vendors and large technology companies will build not only Agents but also developer tools and platforms to compete for the ecosystem; general-purpose Agents therefore look more like a giant’s opportunity. Early-stage investment is relatively concentrated in specialized Agents, while whether vertical industries can produce a US-style SaaS landscape remains unproven, and the lessons from earlier domestic SaaS investments cannot be ignored.
Incomplete Infra is instead generating the most projects: developer platforms, Multi-Agent systems, automated execution, Web testing, data analysis, payments, memory, and search services. MCP heated up suddenly because it attempts to solve the question of how existing tools can be invoked; OpenAI has also released an MCP-compatible Agents SDK.
Specific financings show that capital is already taking positions: Browser Use raised $17M on March 22; Browserbase topped the early-stage Enterprise Tech 30 ranking; and A1Base gives Agents only WhatsApp, Email, and phone numbers, providing the “identity” needed to register, receive verification codes, and access services.

8. Once the Boundaries Are Clear, the Name “Agent” May Disappear

庄明浩’s first deliberately provocative claim is: “If the boundaries are unclear, call it an Agent; once the boundaries are clear, you stop calling it one.” Under that framing, AI search is simply AI search and AI Coding is simply AI programming; once Deep Research clearly takes responsibility for research tasks, people also seem less interested in emphasizing that it is an Agent.
A group-chat complaint captures the conceptual risk precisely: “The last concept that tried to contain this much was the Metaverse.” The more Agent can explain everything, the less likely it is to have a business boundary that can be tested.
庄明浩 says the circulating WeChat Agent roadmap is fake news: IMA belongs to CSIG and is unlikely to become WeChat’s central hub; WeChat data is highly private, making permission calls extremely difficult; WeChat has not even fully rolled out AI search, so it is even less likely to suddenly open the floodgates to Agents.
His view is that WeChat will certainly build an Agent in the future, and Mini Programs could indeed become the Agent’s “hands.” But the implementation and rollout pace will be constrained by privacy, security, and WeChat’s consistently cautious product cadence.

9. Pricing and Automation Value Will Be Tested Before the Concept

“Whether the cat is black or white, a cat that catches mice is a good cat.” Whether a unified entry point will restore the advantage of super apps, whether input must be natural language, whether Agents can gain access to browsers, Mini Programs, and virtual machines, and whether vertical-scenario data still has value all remain unanswered; for now, projects can only be judged by ARR and growth.
庄明浩 believes Agents may not stay at $20 per month forever. He laid out OpenAI’s progression from Free and Plus at $20 per month, to Pro at $200 per month, and the imagined Max at $2,000 per month and “PhD model” at $20,000 per month, arguing that widening capability gaps are opening up pricing room.
Manus brings the cost down to the task level: $39 includes 3,900 credits, $199 includes 19,900 credits, and free users receive 1,000 credits per month. An NBA scoring-efficiency quadrant chart takes about 15 minutes and 200 credits, or $2; a wedding-invitation webpage takes about 25 minutes and $3.6; and a complex interactive application takes about 80 minutes and $9.
Epoch AI’s cautious view points in the same direction: in the “visible future,” letting Agents browse the internet, operate commercial software, and perform standardized white-collar work may be more profitable than training models capable of Nobel-level scientific reasoning. But inference remains expensive, complex environments remain unreliable, and users still need to supervise and adjust—“there is still a lot to do.”

Compass: Meals for the Masses - [Business Breakdowns, EP.211]AI Venture Capitalist: These Tech Predictions Will Change Everything by 2030 w/ Vinod Khosla | #159