E191 | Small, Beautiful Opportunities Are Here: On the New Paradigm in This Wave of AI Agent Evolution
Summary
The 2025 Agent acceleration came from simultaneous advances in code models, RFT, and MCP, with funding and valuations subsequently sending a market signal. Sonnet 3.5 delivered a step change in code generation; OpenAI launched Operator and Deep Research in January and February, respectively; Manus broke out in March, then secured $75M led by Benchmark in May at a $500M valuation. On the day of recording, foreign media also reported that OpenAI was considering a $3B acquisition of Windsurf, while Cursor parent Anysphere raised $900M at a $9B valuation. The deeper meaning of “the real Agent era is here” is that foundational components are beginning to work together—not that a few more chat windows have appeared.
This paradigm shift is not about models becoming more articulate; it is about them learning to adjust their action strategies based on environmental rewards. The Agents of 2023 and 2024 relied on knowledge bases, tools, and human-designed workflows, while the model’s training objective remained conversation. RFT places reasoning models inside environments such as programming and computer operation, allowing them to explore paths, fail, and retry. Tao Fangbo called it “the AlphaGo moment for language models”: reasoning can now be learned from environmental feedback, creating the “big gap” between old and new Agents.
Coding is becoming the universal action layer of the digital world, giving Cursor, Windsurf, and Devin potentially far more upside than traditional IDE plugins. Windsurf’s context engine can recognize code, tests, configuration, the command line, and technical documentation; Devin goes further by bringing the browser, IDE, testing area, and editable “strategy room” and “war room” into the environment. With MCP connections, Cursor could eventually shift from “producing code” to “completing tasks.”
General-purpose Agents could absorb many consumer scenarios that once supported standalone startups, making the defensible opportunities increasingly small and specialized. A single Deep Research could cover academic research and market-research tools; a single Operator could cover grocery shopping, ticket booking, price comparison, and travel planning. The product boundaries controlled by large companies will also be broader than in the App era. Tao Fangbo therefore questions how many service Agents the consumer market can still support, while favoring vertical products in insurance and bidding that have proprietary “weapons arsenals,” evaluation standards, and data environments, as well as “second selves” that express individual identity.
Manus’s advantages begin with productization and controllability; its weaknesses are its general-purpose positioning, data barriers, and broken network effects. Kolento prefers it for broad but shallow research when entering a new field: the process is transparent, users can jump to live and take over at any time, and reports can become interactive websites. Its instruction-based memory is also more behavior-aware than ordinary RAG memory. But “world’s first general AI agent” can attract early adopters while also implying that “generality means having no first-association use case,” and login, permissions, Xiaohongshu, or Facebook data walls are far beyond what even the coding-, GUI-, and command-line “octopus” can penetrate.
The real capability center of an Agent product may not be the prompt but the evaluation layer spanning intent, tool calls, execution, and output. Kolento’s analogy was: “The prompt is the weapon, but evaluation is your crosshair.” Only when every change can be quantified can a Demo become a product that improves continuously. If evaluation against real environments can be upgraded into reproducible rewards, Agents could explore and align themselves; but the discussion noted that RFT may deliver only about a 25% improvement over SFT while multiplying costs, so commercialization must account for the marginal return.
The industry’s tempo will diverge sharply: products are being built “almost frighteningly fast,” while mass adoption and ecosystem formation may still take years. Kolento has shifted from heavyweight PRDs and planning toward faster execution, but also revised “having built more than 200 Agents” to say most were merely chatbots or workflows; Tao Fangbo likewise moved from calling 2025 “the year of agent” to taking a more cautious view. MCP currently looks more like “an AI wrapper around APIs”; certification, privacy, data access, and platform monetization have no unified answers, and opening a technical channel does not mean an Agent can freely enter an environment.
Deep dive
1. Three foundational shifts jointly triggered the 2025 acceleration
Hong Jun laid out a dense timeline: OpenAI launched Operator, which could autonomously use a browser, in January; Deep Research in February; and Manus broke out in March as a “general AI Agent.” By May, Manus had secured $75M led by Benchmark, with the episode putting its valuation at $500M.
During the second recording on May 6, the episode also mentioned foreign media reports that OpenAI was considering acquiring Windsurf for $3B; Cursor parent Anysphere had raised $900M at a $9B valuation. Funding and valuation had begun to serve as market signals of attention toward Agents, especially coding Agents.
Hong Jun summarized the three forces as the step change in coding ability brought by Sonnet 3.5, RFT’s ability to improve performance on specific tasks with limited data, and MCP, proposed by Anthropic in late November 2024 to begin connecting websites and services.
2. New Agents combine general interaction with traditional reinforcement learning
Tao Fangbo first brought the concept back to machine learning: in the reinforcement-learning context, AlphaGo is an Agent because it learns action strategies autonomously from environmental feedback. In the era of large models, the term has expanded to systems that independently complete tasks like humans, are driven by foundation or reasoning models, and have memory and user interfaces.
The large-model Agents of 2023 and 2024 generally had no autonomous learning capability. Their domain knowledge came from developer-configured knowledge bases and workflows; they were more like “scaffolding”—able to call tools, but unable to truly learn how to respond to changes in the environment.
Once reasoning models were combined with reinforcement-learning fine-tuning, Agents began to combine both sides: they could communicate with people and handle complex tasks while exploring paths in specific domains such as law, programming, and computer operation. “Now I really feel that Agents have intelligence.”
3. Environmental feedback fills the biggest gap in the previous generation of Agents
Tao Fangbo’s core explanation was that whether in pretraining, post-training, or RLHF, early language models were trained mainly to talk to people, “not to interact with the environment.” Forcing a model to act as a brain connected to tools does not mean it understands the world those tools inhabit.
He used DeepSeek R1 to illustrate the change. Previously, SFT could give a model problem sets, final answers, and a memorized knowledge base; R1 instead adjusts its reasoning strategy during training and learns how to complete mathematical or programming tasks more effectively. He called it “the AlphaGo moment for language models.”
Companies had already tried using GPT-4 to operate computers in 2023, but the results were poor because the model “had never actually spent time in a computer-operation environment.” Operator puts a reasoning model inside a computer environment for reinforcement learning, making failure, error localization, and action adjustment part of training.
4. A good environment is not a scenario description but an expert’s complete weapons arsenal
A programming environment is not just a browser or code text. It also includes the IDE, testing, deployment, GitHub, open-source projects, and technical communities. A startup’s job is to convert the spaces experts know into an environment an Agent can act in, then add a small amount of effective data and a reward function.
Tao Fangbo’s screening criterion is that any real-world task requiring an expert to possess “a set of weapons” and a specific methodology may support a dedicated Agent environment. Law, medicine, producing Silicon Valley 101, and even the military and weapons operations faced by Palantir all fit this pattern.
This also means that coding Agents need to be retrained for their environments. Tao Fangbo does not believe one large company can build a single model that enables the same Agent to complete every professional task without distinction.
5. Operator turns fixed workflows into dynamic closed loops
Operator opens a browser on the server side and breaks a goal into stages such as searching, comparing prices, selecting, and paying; payment and credit-card steps still retain human verification. Kolento confirmed that he had used it to buy groceries on Instacart, describing it as “very slow, with accuracy also not high enough,” but clearly better than directly driving GPT-4 or the original o1. Kolento also said Operator and Deep Research were both fine-tuned from the o-series models.
It brings basic operations such as clicking, keyboard input, opening a new tab, and Google searches into the reasoning output. Once it generates an action, it pauses its thinking, executes the action, reads the webpage feedback, and continues planning. “Think, execute, receive feedback, continue thinking” becomes a continuous loop.
Hong Jun asked how this differed from a workflow. Tao Fangbo’s answer was that old Agents required developers to “teach them every step by hand”; new Agents learn the broad framework for completing a task. If a website is unavailable, they can switch sites on the fly instead of getting stuck on a predefined path.
Deep Research closes the loop in a similar way, but uses public-information retrieval as its toolset: it revises hypotheses based on search results, adds follow-up searches, and handles conflicts among sources. Both are classified as RFT, with the main differences lying in the action set, dataset, and reward objective.
6. Virtual computers and context engines are becoming Agent infrastructure
Kolento described Scrapybara’s positioning as roughly “a computer for your agent”: instead of running an Agent directly on a personal computer, it launches a secure, scalable virtual desktop through an API within seconds, allowing the Agent to write code, analyze data, browse the web, and take concrete actions. He also said the company had been funded by YC the previous year.
He once used Scrapybara to buy long underwear on Amazon. According to Kolento’s recollection, its website had previously appeared to position finding prospects for sales development representatives as a use case, with the information potentially integrable with Salesforce. What it sells is not hardware but a virtual environment an Agent can operate safely; he speculated that the product later connected to cloud computing as well.
Windsurf’s difference lies in its context engine: it knows where code, files, configuration, and tests are, while also bringing command-line execution and technical documentation from the internet into the action space. Tao Fangbo believes it won not simply because it arrived earlier, but because “its understanding of the environment was more complete than everyone else’s.”
7. Devin expands software engineering into an environment for sustained thinking
Devin never limited the Agent to an IDE. Its interface simultaneously accommodates a browser, code area, user communication area, and testing environment, with the goal of independently completing a longer software-engineering loop rather than merely autocompleting code.
Long tasks exceed the context a model can understand in one pass, so Devin provides a separate notes area that can be repeatedly revised. Tao Fangbo called it a “strategy room, war room”: sometimes the Agent is neither writing code nor searching, but restructuring system designs and action plans in its notes.
The idea made him feel “very frightening,” because the Agent had begun to proactively create a space for thought like a human, then use that space to iterate. Hong Jun also mentioned that the global software-engineering market is “apparently” worth about $4.2T annually.
8. Lower development barriers do not necessarily create more large companies
Kolento’s direct feeling is that “the era has changed”: once you know how to build products, development costs become “almost frighteningly low.” When he encounters needs such as show notes or blog generation, he immediately builds a tool, accumulating about 200 projects across multiple platforms.
He later clarified that most of those projects were “not intelligent enough, and not dynamic enough,” and should more accurately be called chatbots or workflows. Tao Fangbo acknowledged that RFT can reduce dependence on data, capital, and compute, but asked: if everyone can build service Agents, how many Agents are worth rebuilding as standalone products? Deep Research could potentially cover academic research and market research in one shot, while Operator could consolidate 100 consumer needs into one product.
His conclusion is therefore cautious: the individual opportunities controlled by large companies will cover broader demand and attention than in the App era, leaving less room for consumer startups. This is an era better suited to “small and beautiful” products than to repeatedly building service Agents for everyone.
He used public accounts and Douyin to explain a possible reversal. Once platforms are captured by giants, individuals can still become self-media creators or small entrepreneurs through distinctive expression. If AI forms a network in the future, Agents representing individuals and releasing personal identity may create a new market, though the form remains unclear.
9. Manus’s first aha moment came from product visibility
Kolento was among Manus’s first users. His initial aha moment came from the Notion-style UI and the ability to develop temporary webpages into interactive sites that could be permanently published and enter a community. But the website was a visual container for research results, not the primary task.
He more often asks Manus to scrape large amounts of information on biology, neuroscience, or new products to quickly build a panoramic view of a field. His judgment is clear: early outputs contained “a very large amount of information, but the quality was somewhat low,” making Manus better for breadth than depth.
Manus stores memory as instruction pairs—“when the user executes a certain instruction, Manus takes a certain action.” Kolento requires it to ask as many clarifying questions as possible before starting any task, and it continues to follow that rule. By comparison, he believes OpenAI/GPT’s RAG-style user memory is not similarly behavior-aware.
Between Manus and Deep Research, he personally prefers the former. Deep Research has long waits and an invisible process; Manus shows its browsing and coding steps, can be interrupted at any time, and lets the user take over the computer through jump to live, making it more “white box” and traceable.
10. Manus has a powerful general action stack, but harder data walls
Tao Fangbo believes Manus got the combination of coding, GUI, and command line right, creating an intermediate action layer that penetrates different tools “like an octopus” and avoids integrating websites one by one. In real use, however, it remains concentrated in webpages, Deep Research, simple PPTs, and office materials, and has not yet expanded to a broad user base.
The larger constraint may not come from the model or engineering but from data barriers. People can log into different environments using themselves as a “verified token”; AI may not be able to access Xiaohongshu, Facebook, or a user account. Once a critical link is severed, the supposed general capability is cut off.
Sonnet’s significance is not merely higher intelligence but that it was the first to “crack the coding problem.” The digital world was already designed to accommodate code, making coding the universal layer for operating data and environments. But any barrier that coding cannot reach is equally beyond Manus and other Agents.
11. Dependence on models requires products to build their own evaluation flywheel
Kolento wants a product where “the stronger the model gets, the stronger I get, rather than the stronger the model gets, and then I disappear.” He said Manus outsources most of its intelligence to large language models, and speculated that dependence on task understanding, tool calls, and other capabilities makes model upgrades, API fluctuations, latency, costs, and hallucinations uncontrollable variables at the product layer.
He cited Yao Shunyu’s framework: reinforcement learning involves algorithms, environments, and prior knowledge, while evaluation may even matter more than training. He used to focus too heavily on prompt engineering or the latest model; now he cares more about whether system performance can be measured continuously.
“The prompt is the weapon, but evaluation is your crosshair.” Human evaluation is closest to preference but sparse, expensive, and imprecise; code-based evaluation is fast and cheap but struggles with complex semantics; LLM-based evaluation attempts to have models assess their own performance automatically.
For a product like Manus, intent recognition, tool and API calls, execution, and generation should each have independent evaluation logic. Kolento envisioned something like evaluation operations and used frameworks such as Phoenix to abstract templates for summaries, RAG, and code generation. Tao Fangbo went further: only by turning evaluation into reproducible rewards can an Agent optimize itself.
12. RFT’s capability gains must pass a cost test
Hong Jun raised the commercial question: some practitioners believe RFT may outperform SFT by about 25%, but at several times the cost. For a large user base, is it worth paying for that 25% continuously, or should companies start with cheaper SFT?
Tao Fangbo believes the technology industry will gradually migrate toward RFT. Hong Jun said he had heard that the Manus team used SFT internally, and the episode then discussed why on the premise that Manus was not using RFT at the time.
OpenAI did not publicly disclose the RFT process and technical details until the end of December 2024, so Tao Fangbo did not find the use of other methods surprising. More specifically, if a company chooses Claude Sonnet, SFT cannot directly adjust the core model responsible for the chain of thought because Sonnet does not allow fine-tuning. His speculation was that Manus fine-tuned the computer-use model: “They were actually tuning its hands, not its brain.”
13. “General” can capture mindshare while eliminating the reason to use a product
Kolento believes Manus calling itself “world’s first general AI agent” is a double-edged sword. It attracts early adopters and creates hype, but “being able to do anything” does not tell users when to think of you. “Generality means having no first-association use case; it means having no position in the user’s mind.”
A general Agent also raises the cognitive cost of activation. Users must not only learn how to express a need but also understand the boundaries of the capability. Many people simply do not know what to say when faced with an LUI, so Kolento would rather have the Agent clarify proactively and align step by step.
He cited Google Wave as a cautionary example. The 2009 product combined email, IM, collaborative documents, and publishing, but failed because users did not know what it was; its features were later split into Gmail, Docs, and Chat. Manus and Cursor holding community events globally may be intended to consolidate vertical use cases.
At high-risk steps, the experiences of Fellou and Replit better match his expectations: align visually first, then bring a human in the loop for logins, passwords, and similar operations. If Manus fails to identify permission issues in advance, it may burn through credits and still get stuck.
14. General-purpose potential and vertical cash flow offer two ways to place a bet
Tao Fangbo is most bullish on products like Cursor. It has already entered through technical users and is actively connecting data and applications such as Notion through MCP. The interface may eventually split into two modes: one centered on the IDE and code output, and another using code merely as an aid to completing tasks.
Hong Jun asked why not Windsurf. Tao Fangbo acknowledged that the two are close, but Cursor formed its product earlier and has greater market penetration; an independent team may also have more freedom to explore. Regarding the rumored Windsurf acquisition, he reserved judgment, noting that OpenAI might equally inject a general-Agent agenda.
He also did not dismiss Manus. Starting with office research and file handling, then expanding the capability set layer by layer, remains the right path. But “building a general Agent is a very long-term process of continuous accumulation” and cannot be completed through a single breakout moment.
15. The most concrete vertical Agents are already demonstrating measurable value
Kolento cited Vantel, which specializes in workflows for commercial insurance brokers. He said many insurance professionals spend about 80% of their time analyzing policies and entering data; Vantel claims it can save more than 10 hours per week and compare policies hundreds of pages long, with numerous factors, in 5 minutes.
The Sweet Spot serves enterprises, NPOs, NGOs, and SMBs by searching grants, interpreting contracts, and analyzing and assisting with bids and tenders. To get around its 48-hour trial limit, Kolento switched between 8 or 9 email addresses and eventually contacted the CEO directly for membership. That detail says more about product stickiness than an abstract assessment.
Hong Jun and Kolento have both used Gamma for years. Its moat is not just one-click PPT generation, but modular pages, drag-and-drop components, video effects, fine-grained adjustments, and commercial-use image labels built during the pandemic. Tao Fangbo said that even the free version, limited to 10 images, covers most common needs.
16. Products are being built faster, but mass Agent adoption may be slower
The biggest change for Kolento over 2 months was the rise of execution as a priority. Competitors are releasing quickly on both the open-source and closed-source sides, compressing traditional product-manager-style long PRDs and detailed planning; the chain from product need to engineering implementation is becoming shorter. “Everything is accelerating.”
He also corrected what he said in the first recording: most of the roughly 200 projects were “not intelligent enough, and not dynamic enough,” and should more accurately be called chatbots or workflows. The product with the strongest commercialization was a grants pipeline he built in 2023 for midsize and large NGOs and NPOs in East Africa: search, expand answers, ask questions about files, edit documents, and then have a human confirm and submit step by step.
Tao Fangbo’s mindset has cooled in the opposite direction. At the start of the year, he believed it was “the year of agent”; after Manus and MCP appeared, he instead found more problems. Chatbots have already entered the daily lives of the parents’ generation, but whether Agents are built by vertical communities or one or two general products, they remain far from everyday use by ordinary people.
Second Me has been released as an open-source “second self” platform, with the episode putting it at about 92K stars. Practice has exposed a fundamental problem: an Agent is a two-way converter between user needs and environmental action. The environment must be accessible, but people must also have a high-quality channel for expressing preferences and intent.
17. MCP opens the door, but has not solved who is willing to open it
Hong Jun understood MCP as a protocol for multi-Agent collaboration, and Tao Fangbo corrected him on the spot: it is not a protocol between AI systems, but a protocol for an Agent to access external information sources or services.
He expects MCP adoption to be measured in years rather than months or weeks. The first bottleneck may not even be a protocol flaw, but whether platforms are willing to open their data. Interfaces and data are the foundation of commercialization; without industry standards for authentication, privacy protection, and commercial safeguards, technical availability will not automatically produce adoption.
Tao Fangbo’s final characterization was deliberately restrained: MCP is currently more “an AI wrapper around APIs” and has not yet become a truly AI-native entry point. An Agent can knock on the door, but it still cannot pass through freely into another environment and complete an operation.