Summary

The next wave of Agent value lies not only in building more Agents, but in rebuilding an entire stack of software and cloud environments for Agents. 雷磊’s core distinction is that “SaaS is a tool—you have to use it, whereas an Agent delivers the result.” General-purpose Agents and countless vertical Agents will coexist like department stores and neighborhood shops; the latter could even outnumber SaaS products “by thousands of times,” creating demand for new infrastructure spanning identity, payments, code execution and browsers.
Agent infra product forms are shaped by two differences: Agents can run at massive parallel scale, while responsibility for AI behavior remains unclear. The former requires products to move from fixed workflows to real-time feedback loops, allowing Agents to try 100 or even 1,000 paths simultaneously; the latter requires sandboxes to define acceptable boundaries of impact. “The key is designing the best feedback loop,” not making models follow human steps.
E2B and ForeverVM illustrate two product lines for code environments: safety boundaries and low-cost state persistence. E2B uses MicroVMs to provide fast-starting cloud sandboxes and gained visibility through Manus’s adoption; ForeverVM addresses the problem of resuming an Agent after an hour without keeping the environment warm throughout. The real metric is not abstract “security,” but whether the fence can keep impact and resource costs within acceptable bounds without overly constraining capability.
Browserbase has a first-mover advantage, but a runtime-only moat may not be enough. 雷磊 says Browserbase reached a $300M valuation within a year, using the reality that “40% of internet traffic already comes from bots” as its entry point; but product experience, its closed-source nature, and networking, latency and state-management issues in browsers still lead teams such as Manus and Fellou to build in-house. Grasp’s view is that the agentic layer needs to be integrated with the runtime.
Grasp is betting on combining browser runtime with agentic control, leaving developers to add only industry know-how. The product targets four Agent-native problems: cloud execution, headless access, secure login and long-lived stateful sessions; Security Local Login enters passwords and email verification codes locally, while stateful browser sessions let an Agent resume on the original page without consuming resources continuously during the wait. 雷磊’s view: “You have to build the bottom layer and the middle layer together.”
Code and browsers will be the two most important Agent environments, while Browser Use today resembles AI Coding in 2023. 雷磊 explains the adoption threshold as “number of samples × model success rate”: GPT-3.5 pushed coding across the usability line, while web-task samples are larger and success rates remain insufficient today—but model improvements should move Browser Use across the line faster. By his estimate, the global software-development market is roughly $3T-$4T, so a 5% efficiency gain would represent about $150B; sales, recruiting, lead generation and communication on the Web offer similar leverage.
Cloud resources will benefit as the number of Agents rises, but betting on cloud does not mean every cloud vendor wins. 雷磊 calls AI environments “an AWS-scale opportunity,” while invoking Nokia as a reminder that industry expansion does not protect incumbents that fail to transform. He “has no way to judge” when Agents will fully take off, so Grasp may initially target existing but underserved demand in traditional scraping, RPA, automated testing, intelligent sales and marketing.

Deep dive

1. Agents Are Not Rebranded SaaS; They Are a New Service Customer

雷磊 explains why he started Grasp through his own career: he began as a programmer, worked at Google and ByteDance, and built a developer platform at ByteDance. He started a company with his current co-founder in 2022, continued working in developer tools and AI, and recently chose to build “a browser for AI.”
He first rejected the market’s practice of repackaging SaaS as Agent products: “Agents and SaaS are completely different things.” SaaS requires a person to use it, while an Agent should “deliver the result” like a person does; software designed for people therefore cannot simply be handed to Agents unchanged.
On the debate between general-purpose and vertical Agents, he believes the answer is not either-or. The future will look more like a large shopping complex: comprehensive department stores alongside many small shops serving distinct customer groups. As long as a vertical Agent delivers better results in a narrow domain, it can survive for the long term like a shop with a loyal customer base, even without many users.
曲凯 asked whether there could really be hundreds of billions or trillions of Agents. 雷磊 did not confirm a specific number, but believes the total will rise “by thousands of times” relative to SaaS. SaaS is a general-purpose tool, while an Agent is a “storefront” that delivers results and can be sliced more narrowly—even narrower than a Taobao shop.

2. Humans May Assist Agents, but This Is Not a Human–Machine Zero-Sum Game

雷磊 describes the present as Agents serving humans, but the future may shift toward “humans serving Agents”: Agents have higher information bandwidth and can access more knowledge and signals than an individual, with people becoming supporting links in the broader delivery system.
曲凯 kept the central question alive: if humans are no longer the principal actors, who is actually giving the orders—the boss commanding people, or the combined system of Agents and people? 雷磊’s response carried a clear value judgment: “Don’t take yourself too seriously.” If humans helping AI deliver better results is what matters, there is no need to argue over status first.
He also rejects framing humans and Agents as a zero-sum contest. The objective is not for one side to defeat the other, but for groups made up of humans and Agents to keep advancing. The standard remains whether the result moves the world in a positive direction, not who is called the principal actor along the way.
He cites a five-stage AGI framework comprising chatbot, reasoning, Agent and innovative, and says the first 2 stages are already behind us while the industry is currently in the Agent stage. Beyond that, AI may produce innovative outcomes humans never anticipated.

3. Agent Parallelism Will Change the Linear Software Paradigm

The first fundamental difference between humans and Agents is the number of work threads: a person can advance only one task at a time, while an Agent can try 100 approaches simultaneously and select the better path from the results. This is not merely an efficiency gain, but a shift in the working paradigm from linear execution to parallel exploration.
曲凯 connected this to multi-agent collaboration, but 雷磊 corrected him: the 2 are not exactly the same. Humans can collaborate with humans, and Agents with Agents; multi-agent is simply an organizational form. The deeper difference is that even a single Agent executing a task can conduct parallel searches that humans cannot.
He uses algorithms as an analogy: human execution resembles a greedy algorithm, exploring a map incrementally and seeking local optima; an Agent may operate more like dynamic programming, searching for the global optimum. Humans can plan globally first, but actual execution remains constrained by a single thread.
This means handing one SaaS step to a model does not automatically create an Agent product. Software for Agents must accommodate parallel exploration, trial and error, and high-frequency feedback rather than simply preserve the sequential workflow designed for humans.

4. The Best Feedback Loop Matters More Than Pre-Orchestrated Workflows

雷磊 uses coding to illustrate the new paradigm: a person writes the first method, then the second, and finally connects them with logic; an Agent can generate 100 methods first, run them all and receive feedback, then generate 100 or even 1,000 candidates at the next node. The design focus therefore shifts from “how to write sequentially” to “how to provide feedback signals.”
He cites the DeepMind team’s AlphaProof solving olympiad mathematics problems: the system does not need to be taught every step in the way a human would work. It only needs the problem converted into a machine-readable form and a series of feedback signals. Outsiders may not know exactly how it derived the answer, but they can evaluate the result based on whether the problem was ultimately solved.
Grasp therefore evaluates each browser action based on whether it has a positive or negative impact on the final task, then feeds that judgment back into the system as a reward signal. 雷磊 places this within the reinforcement-learning framework, while acknowledging that the industry has consensus only on the direction: “Who can do it, and who can do it better, nobody knows.”
His non-consensus view is that longer context, more knowledge or a better-fitting model are not the primary variables. “The key is designing the best feedback loop.” An Agent product is itself an environment: its value lies in letting AI interact with the outside world, achieve results and continuously improve delivery through rewards.

5. Grounded Signals Determine Whether Agents Can Surpass Human Judgment

雷磊 makes a sharp distinction between real-world feedback and human evaluation, using the term grounded signal. Many large models still rely on RLHF: a person looks at a sequence of proposed actions and judges whether it is “good.” In reality, an Agent should judge itself by whether the task result is complete.
In Browser Use, the reliable signal is not whether a person thinks the clicks look reasonable, but whether the web task was actually completed. The former still treats human preference as the answer; the latter lets the model obtain verifiable results from the outside world.
This is why an Agent environment is more than an execution container: it must also determine whether an operation truly completed and convert the real-world result into a signal usable by the feedback loop. Without that loop, a model may know how to operate a browser but struggle to improve from the outcome of its actions.

6. Unclear Responsibility Makes Sandboxes Critical Agent Infrastructure

The second fundamental difference between humans and Agents is responsibility. Code written by a person runs on that person’s computer, and the person can take responsibility if something goes wrong; if AI-generated code deletes every file, it is unclear who bears the responsibility. The model therefore cannot simply be allowed to act on a complete personal environment by default.
雷磊 avoids summarizing the issue with the vague word “security” and instead emphasizes boundaries: put AI inside a fence and know in advance how much it could affect in the worst case. A fence that is too small suppresses Agent capability; one that is too large exceeds what the user can tolerate. “Finding the right balance is very difficult.”
E2B’s core offering is a code execution environment: isolated sandboxes built with MicroVMs, with extremely fast startup to support parallel trial and error by Agents. In 雷磊’s description, it is a process-level sandbox, faster than a traditional Docker container; the virtual machine is the technical path, while E2B is the solution tailored to the actual need.
Startup speed is not merely an engineering metric. When an Agent runs many candidate paths simultaneously, the faster the environment returns real results, the denser the feedback loop and the faster the optimization and iteration. Security boundaries, performance and learning efficiency are interconnected.

7. Agent Environments Naturally Move to the Cloud, but Cloud Vendors Do Not Equal Finished Environments

雷磊 believes the “only advantage” of local execution is avoiding network latency, while it brings security risks, limited elastic scaling and difficulty running 7×24. If the model is to be sufficiently powerful, it needs to run in the cloud; in that case, placing the model and execution environment together is the more natural architecture, while the client can remain local to display the interface.
曲凯 pressed him on Cursor, which prompts users to close a local port; if a user does not understand code and agrees, the system can indeed run into trouble. 雷磊 acknowledged the risk and observed that engineers are shifting from inspecting what code was generated—similar to conducting code review—to “just let it execute as long as the result is correct,” making isolation and trust more important.
In his view, as Cursor moves from Copilot toward an autonomous Agent, its code-execution architecture will gradually shift from local to cloud. The interface can be returned to users for inspection and trust-building, but users ultimately care about the result rather than which code the model generated.
曲凯 questioned the difference between E2B and the cloud and suggested cloud vendors might eventually build finished environments themselves. 雷磊 used real estate as an analogy: cloud vendors provide utilities and basic computing, while E2B is the contractor that arranges those resources into an environment Agents can actually use. While AI demand is still expanding rapidly, the 2 are primarily partners.

8. E2B’s Value Comes from Agent Pain Points and Boundary Control

E2B’s rise was closely tied to Manus: the application layer needed Agents to generate and execute scripts inside virtual machines, and developers searching for an off-the-shelf solution found that E2B had already solved engineering problems such as fast startup and isolation. They adopted it directly. “A lot of infra catches fire because the application above it catches fire first.”
雷磊 believes the best teams for these products must satisfy 2 conditions: they must be Agent builders who understand the new working model, and deep users of the environment themselves. The same applies to Agent Browsers. Long experience with traditional browser engines matters less than having spent years writing browser scripts and understanding developer experience and real Agent pain points.
On building in-house versus buying, he clearly favors using existing tools. A developer’s value lies in delivering results, not proving that every line of code was written internally. Only when an off-the-shelf solution cannot meet a highly critical requirement should a team build in-house or customize an open-source project.
This leaves a tension in his view. When discussing cloud vendors entering the market, he said that in the AI era, “so-called moats do not exist”; when discussing specific products, he said value created by solving real pain points is itself a moat. He did not reconcile the 2 statements into a single conclusion.

9. ForeverVM Shifts Competition from Security to State and Cost

While E2B emphasizes security, competitor ForeverVM emphasizes “state.” An Agent may run a script, switch away to complete another task, and return an hour later. The environment must resume the prior state seamlessly without remaining warm and consuming resources for no reason.
曲凯 noted that this is the direct consequence of Agent parallelism and back-and-forth task switching. 雷磊 agreed: the pain point may not be obvious in a human continuously writing code, and even years of building code editors may not reveal it. People who build Agents while using code environments are more likely to encounter it.
State persistence ultimately maps to a clear metric: cost. Security itself is difficult to quantify, but teams can measure how much impact the fence permits, how many resources the environment consumes while idle, and whether context is lost on recovery. These metrics determine how many parallel tasks the same Agent can run at low cost.
雷磊 believes the Agent-environment market is large enough to support multiple companies offering different solutions across security, state, language types and specific workloads, without converging prematurely on a single winner.

10. Browserbase’s Entry Point Is Today’s Bot Traffic, Not Distant AGI

雷磊 calls Browserbase a current star and says its valuation reached $300M within a year. The concept is straightforward: move the browser to the cloud and optimize it for AI use cases, such as automatically retrieving website information as context to help models understand and operate webpages.
When Grasp chose the same direction, it followed Zhang Yiming’s idea of “pragmatic romanticism”: imagine countless Agents over the long term, but solve existing demand first. 雷磊 cites the figure that bots already account for 40% of internet traffic; the problems encountered in serving those bots are a sufficiently concrete entry point in themselves.
Asked who drove Browserbase’s rise, he shifted to real-world scenarios such as bot web scraping, automated testing and RPA. Browserbase often addresses situations where bots are not intelligent enough, cannot adapt to changes in webpages, or fail because they lack information about the page.
曲凯 asked why Manus and Fellou did not use Browserbase. 雷磊 speculated that one reason is that Peak was already building in this area; another is that Browserbase’s user experience still has engineering issues and the product is closed-source. E2B is open-source, allowing Manus to continue customizing it.

11. Agent Browsers Must Handle Cloud Execution, Headless Access, Credentials and Long-Lived State

The first difference is the most direct: “AI does not sleep,” so a browser for Agents must run in the cloud. It may also run a large number of tasks in parallel; idle waste that can be ignored on a personal computer quickly becomes a cost problem in the cloud.
The second difference is headless operation. Humans need a frontend, visuals and a mouse; Agents do not necessarily need to operate through vision like humans do. The browser process can run without a frontend interface, directly in the backend. In theory, products built for AI do not need human-style interaction interfaces, although browsers for humans will continue to exist for a long time.
The third difference is login. Users neither want to hand account passwords to a large model nor be called in to type them every time. Grasp’s Security Local Login identifies the login flow and enters the account, password and even email verification code locally. 雷磊 says the entire process is autonomous and never sends the relevant information to the large model.
The fourth difference is long-lived state. After an Agent checks flight tickets on Ctrip, it may go elsewhere to reason and wait for a human decision, then return much later to make the purchase. Grasp’s stateful browser session aims to restore the original page as if “the webpage had never disappeared,” without continuously consuming cloud resources during the wait.

12. Grasp Believes Runtime and the Agentic Layer Must Be Designed Together

雷磊 divides the Agent Browser into 3 layers. The bottom runtime retrieves webpages, executes scripts and renders; the middle agentic layer reads pages, plans steps, generates actions and reasons through tasks; the top Knowledge layer carries vertical know-how for industries such as sales and recruiting.
Browserbase is centered on the runtime and offers the open-source Stagehand framework so developers can build the agentic layer themselves; Playwright has also open-sourced Playwright MCP. 雷磊 acknowledges Browserbase’s clear first-mover advantage but believes that if it only builds the runtime, the moat is indeed not large enough.
His rebuttal is that the agentic layer cannot simply be added independently. Secure local login, long-lived state and feedback loops all require control over the underlying runtime; connecting to Browserbase and building a middle layer leaves some functions impossible to implement. Grasp therefore combines the agentic layer and runtime to solve the common engineering problems as one system.
Grasp targets developers, while Fellou targets consumers; in theory, products like Fellou could be built on Grasp. Developers can bring their own industry knowledge and build vertical Agents faster without taking on the entire underlying engineering burden. 雷磊 expects the product to open “next month.”

13. Code and Browsers Will Become the Two Core Building Blocks of the Agent Ecosystem

雷磊 envisions many environment companies in the future, with each infrastructure product acting like a Lego brick. Developers can combine general-purpose capabilities such as code execution and browser access with their own industry knowledge and feedback mechanisms to build specialized Agents, rather than rebuilding the entire foundation for every company.
He uses ChatGPT’s Deep Research Agent to illustrate the 2 core environments: an o3 model paired with webpage browsing and a Python code executor. Coding handles logic execution, while the Browser interacts with information on the Web; these will therefore be the most important and broadest Agent-environment markets.
The 2 categories will still fragment internally. Browsers may focus on retrieving information or producing information; code environments will have different requirements for scripts, interpreted languages and compiled languages. Further out are environments for mathematical formulas, sensors, embodied intelligence and spatial intelligence interacting with the physical world.
Beyond environments, tools built for humans may also be rebuilt for Agents: whether an Agent needs an independent identity, a phone number that can receive text messages, and its own payment capability. Grasp ultimately chose the browser because the team has years of frontend development experience and deep familiarity with browsers and automation workflows, as well as because the Browser is a critical channel through which Agents reach the real internet.

14. Browser Use Today Resembles the Adoption Threshold for AI Coding in 2023

雷磊 uses a simplified formula to judge when a large model can solve a task: “number of samples multiplied by model success rate.” Once the number of successful outcomes satisfies human needs, the product crosses into the mainstream. Models are probabilistic systems, so the key is whether they can cross an acceptable threshold across large-scale tasks.
GPT-3 in 2022 did not push Coding across the line, while GPT-3.5 broke through the threshold. Browser Use has a larger sample base, but today’s model success rates are still insufficient, so users find it “impractical.” As probability improves, it could become the AI Coding of today—and the process may happen faster.
The commercial opportunity remains far beyond the current scale. By 雷磊’s estimate, the global software-development market is roughly $3T-$4T; a 5% efficiency gain from AI would represent about $150B, while AI Coding today is only worth several billion dollars and approaching $10B.
Browser Use has a broader base: sales, recruiting, communication, presenting results and lead generation all happen on the internet. If AI improves efficiency by 5% in any one link, it could create enormous incremental value. He therefore sees the market as “just getting started.”

15. AI Must Experience the World Directly to Truly Surpass Human Knowledge

雷磊’s most important long-term judgment on Agents is that AI will shift from absorbing human data to “experiencing the world for itself” and using real-world feedback as training data. Large models do not yet truly interact with the world, so he believes the industry may still be at the stage when humans had just acquired intelligence and just learned to make fire.
His Sichuan-cuisine example preserves the full causal chain. If a famous chef evaluates recipes generated by a model, the model can at most approach that chef indefinitely and will inherit the chef’s biases. Actually cooking and tasting the dish, then feeding back whether it tastes good or bad, is feedback from the real world.
True innovation is when the chef sees the recipe and thinks it will taste bad, but the dish turns out to be delicious. 曲凯 added AlphaGo as an example: humanity’s greatest bias may be believing that human knowledge is indispensable to models. 雷磊 calls this line of thinking the “Bitter Lesson.”
The significance of code sandboxes, browsers and physical environments therefore lies not only in executing tasks, but in giving AI first-hand experience. Only when action, outcome and feedback form a closed loop can Agents surpass the ceiling of imitating humans and move toward the innovative stage he describes.

16. AI Environments May Be an AWS-Scale Opportunity, but the Near Term Must Start with Existing Demand

曲凯 brought the discussion back to investment: does a larger Agent population mean cloud-vendor stocks will perform better? 雷磊 believes the broad direction holds—cloud vendors sell resources, the world will consume more compute and generate more data, and cloud resources will become more valuable. But industry growth does not mean every incumbent benefits automatically.
His counterexample is the smartphone market. Today’s market is far larger than it was a decade or two ago, yet Nokia was still eliminated, while companies that originally did not make phones could become new giants. The payoff depends not on the existing category, but on whether a company can transform quickly and keep up with the new resource and product demands created by Agents.
雷磊 calls AI environments “an AWS-scale opportunity.” New companies will expand from the tools and environments closest to Agents, and may eventually build their own browser engines. He sees the relationship as primarily one of mutual cooperation and complementarity.
On when Agents will actually take off, his honest answer is: “There’s no way to judge something that far out.” Commercially, time to market still matters. Grasp may first serve existing workflows such as traditional scraping, RPA, automated testing, intelligent marketing and sales, and then move with those users as they transition to Agents.

AI Scouting Report: AI Agents -vs- Agentic AI, from Imagine AI Live Tucker Carlson: ICE Raids, LA Riots, Strong Economic Data, Politicized Fed, War with Iran?