After OpenClaw, I Only Want to Focus on the Next 3–6 Months | A Conversation with Sheet0 Founder 王文锋

Summary

• 王文锋 frames OpenClaw as validation of a product form, not a fundamental paradigm shift. Manus used reasoning models to unlock deep research; OpenClaw pushes the coding capabilities of the latest models to their limit and makes the long-discussed ideas of proactivity and self-iteration actually work. 曲凯 therefore proposed that “all Agents will ultimately be coding agents.” 王文锋 agreed and went further: every Agent will eventually either wrap a coding agent or be built around one.
• Vertical Agents and traditional SaaS both deserve deep skepticism, because terminal＋skill is converging different use cases on the same execution layer. In Anthropic’s disclosed Claude Code usage data, only 49.7% falls under software engineering; more than half is already office automation, data analysis, marketing, and copywriting. 王文锋 argues that SaaS once scaled expert capabilities through fixed UIs, while Agents directly understand objectives, generate solutions, and iterate—effectively giving every company a context-specific expert.
• Coding Agents may already be capable enough; the real bottleneck is that most people do not know how powerful they are. Before the OpenClaw author became famous, he reportedly made about 1,600 code submissions in a single day. 王文锋 likened that to a year’s output from a 3- or 4-person team; Cursor also used an Agent to build a roughly 3M-line browser in a week. He rates top users at 90 and ordinary users at 10, stressing that the gap is not linear: “Someone who knows how to use it well could be 1,000x more efficient on their own.” Ordinary users may still be doing the wrong things or only simple programming tasks.
• Moving from a million engineers to a billion ordinary users is the most direct product and distribution opportunity in this cycle. 曲凯 estimates that OpenClaw already has more than 2M Agents, though that may include multiple accounts deployed by one person. 王文锋 says Claude Code has reached 1M active users; based on the market share he cited, Codex could have 3–5x as many, putting the two at roughly 5M active users combined. Against his projected billion-user end state, penetration is “not even 1%”; the priority is not building smarter Agents, but flattening the barriers to configuration and management.
• Progress on long-horizon tasks comes from combining model capability with filesystem engineering; true proactivity means letting Agents maintain their own context, identify needs, and manage other Agents. As tasks extend from dozens of steps to hundreds or even thousands, the filesystem makes errors observable and repairable: an Agent can inspect files such as memory and progress, detect deviations, and fix itself. 王文锋 ultimately envisions a meta Agent that identifies problems, configures specialized Agents, completes the task, distills the experience, and then releases them—“AI managing AI.”
• AI may eliminate large numbers of execution roles, but humans will retain taste, tacit understanding, and final sign-off. 王文锋 believes the story of 10 customer-service representatives becoming 1 could spread to engineering teams, with 10 engineers reduced to 2. If a task must be explained in full, humans may already have lost to AI. Those who remain need enough long-term context to “know what someone is thinking with a single click” and take responsibility for the final merge and the “qualified” sign-off.
• Sheet0 is productizing its internal Linear—Slack—GitHub automated development workflow for high-Token users. Once a request enters Linear, AI can build, test, attach screenshots, and submit a PR, turning engineers from Agent commanders into quality inspectors. 王文锋 says individual productivity is up at least 10x year over year; products that once took a year to ship now find “2 months too slow—2 weeks may be enough.” The team spent about $20k on AI coding last month, and its target market is defined not by individual versus enterprise, but by Token consumption.
• This productivity revolution could also contract demand, which is why 王文锋 is only willing to solve problems for the next 3–6 months. If the “one-person unicorn” means removing large numbers of employees, wages, consumption, and upstream corporate demand could create a negative feedback loop rather than an endlessly expanding positive one. His startup strategy has shifted from predicting the end state 5–10 years out to “forecasting as a secondary input, following as the primary one”: “Don’t try to solve bottlenecks people have not encountered yet.”

Deep dive

1. What OpenClaw Proved Was Not a New Entry Point, but a New Form of Coding Agent

• 王文锋 sees no fundamental difference between this wave and the one a year ago. Neither Manus nor OpenClaw began by inventing a new theory; each demonstrated a product form that no one had previously confirmed could work. Manus unlocked deep research through the reasoning capabilities of models, while OpenClaw unlocked coding.

• 王文锋 rejects the explanation that OpenClaw became popular mainly because it was easy to use inside IM platforms such as Telegram and Discord. The entry point is superficial. The core is that it “really pushes the coding capabilities of the latest models to the limit,” showing that things once thought impossible to run can now run.

• The industry spent a year repeatedly discussing proactivity, self-iteration, and self-evolution without a representative product. OpenClaw’s significance is that it turned those abstractions into something tangible. 曲凯 used that to revisit 王文锋’s view from a year ago: “AI coding is the retail end of large models.”

2. All Agents May Converge Into Shells Around Coding Agents

• 王文锋’s new conclusion goes further: “All Agents will ultimately become shells around a coding agent.” OpenClaw includes components such as memory, Slack, Discord, and Telegram, but its intelligence core is still the deliberately minimalist Pi coding agent.

• Last year’s discussion assumed that every vertical use case would need its own RL environment, interaction model, and customized strategy. Today, terminal＋skill largely unifies those environments. Vertical know-how that was previously difficult to replicate can now be implemented by pairing a coding agent with the relevant skill.

• Manus was mocked as an API wrapper, but 王文锋 sees it as “the ultimate wrapper in the model API era.” OpenClaw is the “best” wrapper of the coding-agent era. Whether it has reached the ultimate form is still unanswered, but “who can build the best shell around a coding agent” will be a major opportunity this year.

• 曲凯 then asked which path would win: general-purpose platforms or vertical Agents. 王文锋 said he was “reluctant to admit it,” but the conclusion is clear: the vertical-Agent route now “needs to be viewed with deep skepticism.” Vertical positioning can sometimes be a psychological safe harbor for avoiding direct competition with general-purpose Agents.

3. Agents Break Through SaaS by Redistributing Expert Capability

• Anthropic’s usage report offers an immediate counterexample: only 49.7% of Claude Code use cases are software engineering, while more than 50% already involve office automation, data analysis, marketing, and copywriting. “This isn’t something we are projecting; in reality, Claude Code can already do everything.”

• 王文锋 breaks down SaaS’s value this way: software was the first mechanism that allowed expert experience to be replicated at scale without years of training. The UI of vertical SaaS essentially hard-coded an industry expert’s workflow into forms, dashboards, and records. Lower development costs from AI coding are only the surface-level disruption; the real substitution comes from changing the vehicle that carries expert capability.

• To B software historically emphasized best practices because software could not handle each company’s specific reality; it could only force users to converge on a standard process. For AI, executing a best practice and executing a nonstandard process are not fundamentally different. An Agent can understand the objective, formulate a plan, and adapt it to the company’s context.

• 曲凯 offered a useful rebuttal: if coding were the only SaaS moat, cheaper Chinese engineering companies would have displaced U.S. SaaS providers long ago, but that has not happened. The real question is whether skills can carry complex know-how. 王文锋 still leans toward yes, arguing that today’s models have reached a roughly human level: they can learn tools and execute workflows from written instructions.

4. The Capability Ceiling Has Been Crossed; the Bottleneck Is That People Do Not Know How to Use It

• 王文锋 believes the debate over whether Agents can complete complex work is already behind the facts. Before the OpenClaw author’s product went viral in January, he reportedly made as many as 1,600 commits in a day. By 王文锋’s rough comparison, that equaled the output of a 3- or 4-person technical team over a year.

• Another benchmark came from Cursor, which reportedly used an Agent to build a browser with roughly 3M lines of code in a week. The program also mentioned Anthropic’s official use of an Agent to build a C compiler intermittently. Few real-world software projects are more complex than these, suggesting that long-horizon task capability has “reached this boundary.”

• The real discontinuity is on the user side. If a top user consuming $20k–$30k of Token per month scores 90, an ordinary user may still score only 10. The gap is not 8x; it may show up as a 1,000x productivity difference, with the latter user perhaps still doing the wrong thing or working only on simple programming tasks.

5. The Million-Engineer Market Is Only the Starting Point for Billion-User Infrastructure

• 曲凯 estimates that OpenClaw already has more than 2M Agents, though that may include multiple accounts deployed by one person. Manus has several hundred thousand users, and Cursor is in the million-user range. 王文锋 focuses on a different figure: Claude Code has roughly 1M active users; based on the market share he cited, Codex may have 3–5x as many, putting the two at “probably” about 5M active users combined.

• These users are still primarily engineers. 王文锋 believes coding agents will ultimately reach 1B people and become infrastructure. Against that end state, current penetration is “not even 1%”; ordinary operations, sales, and product employees are the large unopened market.

• Many teams today simply provide VMs and hosting for OpenClaw, still serving engineers who can configure it themselves. Even when 王文锋’s team actively encourages nontechnical colleagues to use it, it finds that they “really cannot configure this thing properly.” Helping ordinary users use it well is the real product challenge for this year’s wrappers and distributions.

• OpenClaw has been compared with the Linux kernel, suggesting that Red Hat-, CentOS-, or Ubuntu-style distributions may emerge next. 王文锋 cautions against copying that model mechanically. He believes at least 1,000 teams globally are already working on this space; the value will not come merely from hosting another instance, but from finding a complete usage model for ordinary people.

6. The Filesystem Turns Long-Horizon Tasks From Fragile State Machines Into Repairable Processes

• The most intuitive measure of a long-horizon task is the step: reading a file, accessing an API, and calling a tool each count as one step. Most tasks today still cluster around dozens of steps. 王文锋 expects that to expand to hundreds or even thousands this year, with the key capability being continuous decomposition and progress maintenance.

• 曲凯 revisited the mathematical problem raised a year ago: even with a 90% success rate at each step, the probability of overall success falls rapidly when the steps are chained together. The system may also spin in a loop or enter an endless cycle. 王文锋 believes this problem “should be solved,” because Agents no longer depend on an irreversible in-memory state machine.

• When memory, progress, and task state are written to files, 王文锋 argues that an Agent can reread the situation, identify errors in earlier steps or records, modify them directly, and continue. The goal is not to make every step permanently error-free, but to make errors observable, reversible, and self-repairable.

• Model capability is only one part of the solution. The other is an engineering methodology built around APIs, computer access, and the filesystem. Last year’s goal was for humans to control context precisely; today the most effective approach is to give the Agent a computer and a folder and let it organize its own data and logic: “less structure, more intelligence.”

7. True Proactivity Means Letting AI Identify Needs and Manage AI

• The most basic form of proactivity is repetitive execution: sending yesterday’s work summary every day at 8 a.m., or reviewing unfinished items in group chats at 10 p.m. It does not require a new prompt each day and can optimize the workflow as it runs, but 曲凯 notes that this remains “proactivity triggered passively.”

• A higher level of proactivity comes from long-term context. Once an Agent understands a person’s business, personality, role, and work, it can identify problems and propose solutions on its own, asking “like a colleague” whether the user wants to handle something a certain way. This corresponds to 王文锋’s reference point: “context, not control.” The goal is for the Agent to explore, reflect, summarize, and iterate independently.

• 王文锋 describes the product paradigm as a shift from chat’s “human to AI” toward OpenClaw-style “AI managing AI.” Once humans no longer have to initiate every request, he envisions a meta Agent that creates specialized Agents, assigns them specific problems, records successful approaches, and releases them when the task is complete.

• This form of true proactivity is not yet available out of the box. To understand a user, an Agent first needs connections, permissions, and a period of calibration. A company-specific skill can accumulate experience, but an Agent optimized for one company’s codebase and processes will often have limited value when moved to another. The capability is portable; the local context still has to be rebuilt.

8. Engineers Are Moving From Agent Commanders to Final Quality Inspectors

• Sheet0’s internal workflow originally looked like this: user feedback entered Linear, the daily meeting distributed it to engineers, and engineers drove Claude Code to develop, test, submit PRs, and merge them. Once Linear, Slack, and GitHub permissions were connected, AI could receive a task, develop and test directly, attach results and front-end screenshots, and submit the PR.

• The roles have therefore reversed. Engineers used to command AI; now AI monitors engineers and completes most of the work, leaving engineers to inspect the PR. 王文锋 describes the shift as moving from producer to “final quality inspector,” and says the workflow went from initial experiment to fully functioning in just 1 week.

• 曲凯 asked why the final step could not also be handed to an AI such as CodeRabbit. 王文锋 distinguishes code review from release responsibility. The remaining human action may be nothing more than a quick scan and a click on merge, but even a production line with a 99% yield still needs someone to issue the certificate of quality. That click carries final accountability.

• 王文锋 says his personal productivity is up at least 10x from the same period last year. Products that once took a year to ship now find “2 months too slow—2 weeks may be enough.” Production capacity is no longer the bottleneck; what is scarce is deciding what to build and what good looks like. He therefore reverses the popular slogan: “idea is not cheap.”

9. Agent Harness Is Closer to This Year’s Product Gap Than a New Model

• 王文锋 is most excited about Agent Harness this year. The concept emerged around September last year in small Silicon Valley discussions. Context engineering made the industry realize that context determines success, but did not show people what context engineering actually looks like. Harness is like the saddle needed to ride a powerful horse: it constrains and directs an Agent that is already highly capable.

• It is neither traditional infra aimed only at developers nor SaaS with a clear infrastructure–application–end-user hierarchy. Harness is closer to scaffolding for end users. It must combine a team’s business, personnel, and working methods to create a local environment in which each company can get Agents running quickly.

• 王文锋 says he has not yet seen a particularly clear opportunity in new infra, at least for now. He mentions products such as Obsidian as possible context-management tools, but is unsure whether they can become businesses. He agrees with Andrej Karpathy’s emphasis on the CLI: what models or researchers lack may not be MCP or APIs, but a better Command Line Interface for getting work done. OpenClaw itself relies heavily on the CLI. Permission management is another possible infra opportunity.

• Backend capability historically lagged the front end because the front end provides immediate visual feedback, while databases and in-memory state are invisible to models. Sheet0 has opened dedicated state interfaces to AI and can grant read-only database access, allowing it to check expectations, identify errors, and iterate. Permission management may be an infra opportunity, but standardizing it into a product remains difficult.

10. Sheet0 Is Moving to AI Coding’s Third Stage: AI Driving AI to Write Code

• Sheet0’s new version productizes its internal Linear workflow, shifting from building a “smarter Agent” to building an AI that manages AI. 王文锋’s judgment is close to absolute: “After careful configuration, there is nothing a coding agent cannot do.” The real difficulty is configuration, orchestration, and management.

• He divides AI coding into 3 stages. Stage 1 is Copilot-style completion. Stage 2 is the end-to-end execution of Claude Code and other coding agents, still driven by proactive prompts from programmers. Stage 3 is AI clarifying business requirements, generating tasks, and driving coding agents, with programmers performing only the final acceptance.

• 曲凯 sees the choice as attractive partly because the team itself is a user, creating a positive feedback loop between development and product validation. 王文锋 describes the mission as a form of “equalization”: rather than making a 90-point expert marginally better, first bring people who currently operate at 10 points above 60.

• In the initial user profile, roughly half are founders, one-quarter are “super product managers,” and one-quarter are “super builders.” The shared traits are not conventional titles, but an abundance of ideas, clear thinking, strong ownership, and the ability to work across design, product, and engineering while handling multiple domains at once.

11. Token Consumption Will Redraw Customer Segmentation and Organizational Management

• Sheet0 spent about $20k on AI coding last month. 王文锋 believes cultivating high-level AI users is possible, provided organizations offer “unlimited Token consumption quotas”: get everyone up to speed first, then measure usage precisely.

• Customer segmentation in the Agent era may no longer be To Enterprise, To SMB, or individual users. It may instead be based on Token consumption. Serving an account that uses more than 100M Tokens would be analogous to serving a Fortune 500 company in the SaaS era. Whether that account is one person, a team, or a company becomes less important to the business model.

• His math is straightforward: if one user can contribute $100k per year, serving 1,000 users produces $100M ARR. It was historically difficult to collect $100k from one person; now that amount may simply be an engineer budget being displaced, while a Silicon Valley engineer can cost $150k.

• Organizations are already reorganizing around this new means of production. Some 10-year-old companies are forming Agent task forces from the top 10 Token consumers across the company, reporting directly to the CEO or CTO. Managers must check whether “consumption matches output”; if it does not, “solve the problem or solve the person.”

12. A One-Person Unicorn Could Also Shrink Total Demand, So the Bet Is Only on the Next 3–6 Months

• Sheet0 currently has only 7 people, and 王文锋 says its output is roughly equivalent to the old output of 30–50 people. By traditional standards, the team should already have 10-plus to 20 people. He expects the customer-service pattern of shrinking from 10 people to 1 to spread to R&D: 10 engineers may leave only 2, with the remaining value concentrated in taste, tacit understanding, and final approval.

• 曲凯 asked whether the market would need so many products if everyone could become a “one-person unicorn” with AI. 王文锋’s answer is relatively pessimistic. Layoffs reduce wages and consumption; weaker demand from consumer-facing companies can then spread upstream, creating a negative feedback loop. In some sense, AI is “taking out the people in the middle,” which could also put pressure on social stability.

• That uncertainty has shifted 王文锋 from solving problems 5–10 years out last year to considering only the next 3–6 months. Sheet0 originally aimed for 100% explainability and 100% accuracy; the end-state goal is 100% trust. The direction is right but premature: models cannot yet achieve it, and users have not actually stopped using them because of insufficient accuracy.

• 王文锋’s startup discipline is: “Don’t try to solve bottlenecks people have not encountered yet.” Last year, when the team debated whether to follow Manus and build Genspark, only 1.5 people supported the idea. He now puts Genspark at “apparently around $150M,” without specifying the metric. The lesson is not to copy it, but to set aside differentiation ego, make “forecasting secondary and following primary,” and move only half a step beyond a real bottleneck.