Agent Era Day 500: Stop Investing in GUI-Thinking Software

Summary

• The core call: We may no longer want to invest in software built around GUI thinking. 天杰 at ZhenFund’s framework is that GUI is “a patch for human cognitive deficits”: human context is limited, people can remember at most 7 parallel words, and Notion, Figma and Slack are all products optimized around that limitation; “once something no longer needs human participation, it no longer needs to serve human shortcomings.” The signal: since the start of this year, a growing number of companies have proactively opened MCP/CLI interfaces to agents, including Sentry, Supabase, MongoDB, Feishu and Google Suite—a trend investors need to take seriously.
• The 500-day scorecard shows a clear split: models have beaten expectations, while productization has fallen far short. The products on the market still require users to continuously pour in attention; 归藏 takes the opposite view, calling the overall outcome “beyond expectations to an unimaginable degree”—someone with only basic front-end skills now maintains a codebase approaching several hundred thousand lines, has automated his entire workflow, and has even handed social-media writing to AI: “This insistence is meaningless.”
• Opening the platform layer to CLI is a prisoner’s dilemma. WeChat, Xiaohongshu, Meituan, Didi and Ctrip have not opened theirs because “once you launch a CLI, the user’s entry point is no longer with me,” leaving them as headless service providers. 天杰 expects Meituan, Didi and Ctrip not to open everything to everyone, but to sign agent-to-agent bilateral special protocols with WeChat, iPhone and OPPO/VIVO—“Meituan, Didi and Ctrip are defecting” is already happening. WeChat’s defensive window is narrowing: Douyin Chat DAU is already 100M, and “where the content is, chat will be.”
• Skill is an underappreciated commercial layer. No agent has reached Claude Code’s dominance, but a skill can—cross-agent distribution can drive installs, as seen in 归藏’s PPT skill being incorporated into Open Design, which has 66K stars. Consumer monetization remains difficult: “I’ve already paid for the tokens—why should I pay you again?” 归藏 expects models eventually to absorb skills, but skills have already become a consensus layer and, by AI’s definition of long term—6 months to 1 year—they will continue to exist.
• China’s Claude Code will most likely come from a model vendor. 归藏 says domestic models still lag the top 2 on long-horizon tasks and agentic loops, while leading developers would rather pay a premium for a strong model: a weak model “doesn’t just hurt your efficiency; it also damages your codebase.”
• This is still the great infrastructure era, and downstream applications have not reached their best moment. 天杰 compares the moment with the 99 Super Bowl dot-com ads: spending heavily on user acquisition was wrong then, while refusing to spend in the mobile-internet era was also wrong—timing is everything. The current priority is making tokens smarter and cheaper; downstream applications may have a 30-year runway, but the cycle could be shorter, with products appearing or taking shape within 5 years. Token prices move with energy, memory, model capability and capital markets: “The $200 Claude Max you buy today and the Claude Max you buy 3 months from now have completely different quotas.”
• The next Douyin may still be Douyin—just with 10x the output. Front-end formats have converged, but 100x-to-1,000x gains in production efficiency will move content creation from household pig raising to industrial farms, creating major leaders: “The next video will still have the form of video; only the way it is produced will be completely different.”

Deep dive

1. The 500-Day Review: Models Beat Expectations, Productization Fell Far Short

• Start with the timeline: roughly 500 days have passed since the podcast’s opening episode at the start of the “Year of AI Agents.” Day 100 brought Manus, with Claude Sonnet 3.5 providing the foundation for building agents; day 300 brought Claude Code; day 400 brought OpenClaw. The concepts and technology are still proliferating.
• 天杰’s scorecard is blunt: “Model capability has exceeded my expectations, but productization has fallen far short.” Products on the market are still built around humans as users; “people still have to keep pouring their attention into a product.” He has yet to find a satisfactory product for information consumption.
• 归藏 sits at the other extreme, calling the overall outcome “beyond expectations to an unimaginable degree.” “I absolutely couldn’t have imagined that I would be able to maintain a codebase of nearly several hundred thousand lines and provide services to so many people.” For someone with only basic front-end skills, that would previously have been impossible. Outside of in-person interaction, his entire workflow has entered an AI-automation system.
• His reversal on AI writing is worth noting. He once said he “absolutely would not use AI to write content,” but now finds that “this insistence is meaningless.” The essence of public-account writing is to convey information with maximum efficiency and minimal friction, which will inevitably move toward AI. Literary work is different: AI still cannot reach that bar, so humans remain necessary.

2. Context Is the Constant; Concepts Keep Losing Their Magic

• 归藏’s one-line summary of the past 500 days: “Everything eventually comes back to context.” That applies both to the model’s own context and to context management on the agent side; context is becoming more important across the stack.
• 天杰 offers the counterpoint: what never changes is the steady disenchantment of new concepts. The old joke is that “any technology that has not yet been mastered can be called AI.” Once it is stable and embedded in workflows at scale, “we’ll say, oh, that’s just software.” What also remains constant is that model progress will continue unlocking “technological wonders we had not previously imagined.”

3. GUI Is a Patch for Human Cognitive Deficits

• Three months after the article was published, 天杰’s core logic is unchanged, but he stresses that “thinking” is the qualifier. GUI itself is not going away—OpenClaw and Manus, the breakout products, are products of good GUI design. What is receding is the assumption that humans should remain the protagonists of efficiency and production tasks. Once agents outperform parts of the human workforce, GUI becomes “an additional burden” for agent-first workflows.
• The key line follows from GUI’s design premise: it is optimized around human attention, on the assumption that “the way humans interact with information is flawed.” Context is limited; people can remember “at most 7 words in parallel,” and the cognitive burden rises sharply when they reach the 8th or 9th. Notion, Figma and Slack all deliver friendly experiences by compensating for that deficit. The conclusion: “Once something no longer needs human participation, it no longer needs to serve human shortcomings.”
• 归藏’s daily experience provides the practical proof: “A lot of GUI is actually anti-human, but we had no alternative at the time. Now we do.” AI can absorb the complexity of the CLI while users focus only on the result—“I need to know my meetings for next week right now.” He uses FFmpeg to edit video and added live-photo layout capabilities to his own skill. Even KFC and Luckin have launched CLI interfaces so AI can place an order for you. Offline brands want to “break free from the influence and extraction of the old channels” in the AI era.

4. Notion’s Dilemma: Don’t Bolt Agents onto a GUI at Its Peak

• From the host’s perspective as a 10-year user, the GUI-thinking stack grew more complex from Evernote to Quip to Notion. Notion’s block-based information architecture “reached, in a certain sense, the peak of GUI thinking.” Its recent versions, however, have been “forcing the concepts of agents and AI onto a GUI product that had already been polished to near perfection,” including 5 rounds of changes to accommodate the AI sidebar.
• 天杰’s judgment is unequivocal: “Those products are destroying the value network of their old users rather than opening up a new pool of value.” Force-fitting the 2 products together “only creates conflict between them and makes both groups uncomfortable.”
• 归藏’s proposal to Notion’s product VP: accommodate existing users and leave their familiar environment untouched; give new users a separate, AI-first entry point where the product can be headless. B2B software “shouldn’t be changed frequently, because it is already very complex.”
• The general rule for founders is straightforward: “The most important job of a PM is to define the problem—don’t solve a problem from the previous era that has already been solved well.” The strongest objection to the article—that Douyin’s phone-based interaction is already optimized to the limit—is answered by the same principle: agents are solving a different problem.

5. The Death of Gmail: Sacrificing Its Capabilities to Codex, but GUI Is Not Quite Dead

• 天杰’s personal Gmail experience is the clearest case study. After subscribing to hundreds of email sources, his inbox collapsed. He spent 3 hours every week manually maintaining filters and tags until one misconfigured filter sent every email into a random label. “This is a major source of friction in the way humans use GUI”; the burden of managing the format had exceeded the value of consuming the information.
• He now authorizes Codex to access Gmail. A scheduled task reads the previous 24 hours of email every day in his preferred way and reports only what he cannot miss. “In a certain sense, Gmail’s GUI is dead”—Gmail “sacrificed the capabilities of mail inside itself to Codex,” leaving the user isolated from the service by an agent.
• That leads to a fork in the road for headless companies: “It depends on how broad their vision is.” Companies that fail to transform “will die, reduced in the end to a database.” But if a company integrates exceptionally well with agents, “it will become the default underlying execution tool for everyone,” potentially with a new business model.
• 归藏 immediately adds a guardrail: “GUI is still very important.” Codex won users precisely through its experience—an in-chat video player, mixed text-and-image layouts, and file paths that open directly when clicked. Those are the friction points where other agents simply return a path.

6. CLI 101 and the Big Tech Opening Game

• 归藏’s definition for ordinary users: GUI means clicking buttons to get a visual presentation; CLI is pure-text commands, “similar to pseudocode.” The core reason people could not use CLI in the past was that they could not remember it. Even the author of FFmpeg could not memorize hundreds or thousands of commands and thousands of pages of documentation. “But AI can remember them, so suddenly this becomes very comfortable.” The biggest problem disappears.
• Who should open a CLI but has not? “Most of the monopoly-like software we use every day has not launched its own CLI”—WeChat, Xiaohongshu, Meituan, Didi and Ctrip. The motivation is obvious: “Once you launch a CLI, it feels like the user’s entry point is no longer with me.” The company becomes a headless service provider, while others can use the CLI to extract its data. That is why 天杰 says he admires companies that remain open despite already being mainstream, such as Google and Feishu.
• 天杰 adds a more measured view. For big tech, tool products are not necessarily profit centers, and the cost of opening access is low. His actual prediction is that Meituan, Didi and Ctrip will not open everything to everyone; they will sign limited, bilateral agent protocols with WeChat, iPhone and OPPO/VIVO. They announced their integration with the WeChat agent 2 days earlier, prompting the headline-style framing that “Meituan, Didi and Ctrip are defecting.” Companies serving the physical world have little reason to lock down the “head” online.
• At the most basic level, every company exists to exchange value with the outside world. A CLI is simply “an interface for exchanging demand with an agent.” “If you believe the value your company delivers can reach users better through AI, you should open your interface.” Luckin’s mission is simply to get caffeine into you faster. The secondary benefit is exposure: a novelty news story like Jingu Yuan Dumpling House. It is like the old headline, “We now accept Bitcoin”—few people may actually pay that way, but everyone sees the news.

7. The WeChat CLI Wish: Where Content Goes, Chat Follows

• The answer to the wish list is unanimous: “Of course it’s WeChat. It’s so painful.” When everything else can be automated by AI, communication with people, customers and counterparties still requires the painfully manual GUI process of clicking in and typing a reply. But nobody believes they can persuade 张小龙. 天杰 no longer expects legacy services to transform; he is waiting for AI-first services to emerge. He already uses Manus to find 鸟岛 rather than searching Xiaohongshu.
• 天杰 sees a possible break in the stalemate: if people could add one another on Doubao and let agents read their chat histories, “I would genuinely have some incentive to move part of my communication to Doubao.” At that point, WeChat might be forced to open up defensively.
• The evidence is already emerging. 归藏 observes that young people have started chatting on Douyin, where chat DAU has reached 100M. 扎克伯格 once acknowledged internally that he had underestimated TikTok: he thought it was merely a content platform, but did not anticipate that “when you see content and share it with someone, a conversation starts automatically,” putting WhatsApp under pressure. The logic closes cleanly: “Where I consume content is where I will chat.” If WeChat blocks Douyin links, communication will gradually move elsewhere. “If one day Doubao, WeChat and Feishu are fully connected, that would be pretty frightening.”

8. The Skill Economy: Distilling Taste, and the Monetization Trap

• 归藏 uses only skills he wrote himself, with the sole exception of the official Feishu skill series. His top recommendation is his own PPT skill. Monetization still comes through sponsorships and Token Grant: “There is no way to charge consumer users.” The user mindset is, “I’ve already paid for the tokens—why should I pay you again?” Only after seeing a high-quality result and comparing it with alternatives might users come back and pay.
• The process of generating a skill contains its own insight. The first version of the PPT skill consisted of 1 sentence. AI used his memory and past projects to “dig out some style code” and assemble a theme that matched his taste. “Through the process of using context and your taste, it distills your taste into a result.” Automatic skill summarization works too—such as the prompt shared by the Codex author for extracting 5 or 6 skills from chat histories—but “an AI summary may not meet your requirements, because it is guessing.” Humans still have to adjust it.
• 天杰’s daily use cases show the profile: titling podcasts, finding memorable lines and writing summaries; summarizing meetings into Notion using different templates for friend conversations, angel investments and podcast pre-interviews; and “grill me”—when he cannot think through a problem, asking AI to question him from every angle and “squeeze” out the ideas in his head to help him make a decision.
• Will skills persist? 归藏’s answer is that “ultimately, everything will be absorbed by the model,” just as GPT Image 2.0 compressed the complex prompts of the Banana era into 1 sentence. But “in AI, 6 months or 1 year counts as long term.” Skills have already become a consensus and will last longer. 天杰’s end-state is branding rather than taxonomy: ordinary people cannot perceive the nuance between 2 skills—“your grandmother is not going to learn what a skill is.” They will trust the agent brand that completes the task. “When one company helps you complete every task, that company itself becomes synonymous with the task,” just as nobody says “PageRank search”; they just Google it. The fact that people are still debating definitions shows the product has not reached its terminal form.

9. Who Will Build China’s Claude Code? Skill’s Dominance Is Being Universally Underestimated

• 归藏’s chain of reasoning starts with China’s slower overall agent progress. One reason is that “our models really are not as good as the top 2 on long-horizon tasks and agentic loops.” What puzzles him is that if “software is not a moat and anything you see can be obtained,” product experiences should be able to catch up with the leaders faster. His conclusion: “A dominant agent will most likely come from a model vendor,” with a leading model pulling the product forward. Top developers would rather pay a premium because a weak model “doesn’t just hurt your efficiency; it also damages your codebase.”
• He then reverses the frame: no agent may reach Claude Code’s level of dominance, but “a skill can.” A skill can be installed across a fragmented agent landscape, and once its installed base is large enough, “it must have commercial value.” The problem is that “nobody sees the commercial value of skills”; everyone is instead frantically building their own agent. At launch, platforms simply crawl GitHub, make no differentiated optimizations and do not co-build with authors.
• That is why he welcomes skill listings on Xiaohongshu. His PPT skill works well on GitHub, but “there is no way to show my results or see reviews.” Xiaohongshu supplies the missing loop of showcasing, tagging and comment feedback. There is no authoritative installed-base ranking, so heat is the only proxy. Open Design, which incorporated his PPT skill, has 66K stars; the estimate is pieced together with channels such as the skills.sh marketplace.
• The ecosystem remains crude. Skills cannot push updates; they can only notify users. One workaround is to put an update-check instruction at the top of skill.md, but it takes effect only when triggered—“the old and new versions are a collapsed state for the user, Schrödinger’s cat.” When to build a CLI, a skill or an MCP “is currently a very chaotic question.” Gary Tan packaged YC Office Hour transcripts into a skill, while Lenny packaged podcast interviews into one for users to query. 天杰 sees this as “information equalization.”

10. The Agent Economy: The Great Infrastructure Era, with Floating Token Prices

• For the To-Agent infrastructure stack—sandboxes, memory systems, agent databases, payments and agent networks—天杰’s advice is to start with open-source frameworks such as Pydantic AI and AI SDK rather than building from scratch. He has encountered many projects that give agents WhatsApp accounts, phone numbers or email addresses, but has not used them.
• 天杰’s timing framework is the backbone of the section. During the 99 internet bubble, dot-com companies bought Super Bowl ads but could not retain the users they acquired. Companies scarred by that destruction then became too afraid to spend on acquisition during the mobile-internet era—“in hindsight, that was also wrong.” ChatGPT launched only 500 days ago. “We are still in the great infrastructure era”; making tokens smarter and cheaper remains the main theme. Downstream applications could have a 30-year runway, but the cycle could be much shorter, with products appearing or emerging within 5 years. “Now may not be the best time to build them.”
• One outcome 天杰 considers certain is that “replacing humans in certain exchanges of value will necessarily create a new network that exists outside the human network.” That includes finding 小岛 and screening for matches at the level of a finished work. “It is not that society lacks a suitable person for him; it is that this person’s information has not been made public to this network.” But such a network “may not be able to exist while tokens are still relatively expensive.”
• The reality check for the token economy is that prices move with energy and electricity, model quality, demand and capital markets. “The $200 Claude Max you buy today and the Claude Max you buy 3 months from now have completely different quotas.” Rising memory prices push up the cost of long-context tokens, while greater production will eventually bring them back down. SpaceX, as an unexpected beneficiary, captures the broader point: “I don’t need to train models. I just need to sell compute and infrastructure, and I can still get a very high valuation.”

11. OpenClaw at 100 Days: Forcing Consensus, Proving CLI Is Hard to Use

• The hype faded with embarrassing speed. Mentioning “raising shrimp” now feels “like telling a friend a really tacky old joke,” roughly from the same era as “Aoligei.” But 归藏 believes OpenClaw left something substantial behind: it “forced everyone to complete a cognitive shift in an extremely brutal way.” By January, the industry already understood that Claude Code and Manus “were completely different things,” but investors, founders and users had not made the transition. OpenClaw completed that lesson and popularized the concept of skills, which is why skill installs began to take off.
• 天杰’s reverse conclusion is equally important: “OpenClaw at least proved that CLI is still difficult for most users.” Users want to maintain a context inside a familiar IM product—WhatsApp, Slack and eventually WeChat—in a conversation that feels like talking to a person, not across multiple threads. “For the first time, it made you feel that you were not talking to a tool but to a person.” The failure mode was straightforward: after a difficult installation and an inexplicable crash, “you realize you have no motivation to restart it.”

12. Token Grant and the Next 500 Days: A 10x-Productivity Douyin

• ZhenFund and Crossroads’ Token Grant gives 0-to-1 creators RMB50K in pure sponsorship with no equity attached. The logic is that the missing ingredient in starting a company today is not necessarily money—“AI can help you do so many things”—but “if you want to build something with AI today, you need tokens.” The program has supported 归藏’s CodePilot, an open, local agent that keeps the full harness of skills, memory and CLI on-device; it can serve as a shell around Claude Code or Codex and connect to any cost-effective third-party API. It has 6K GitHub stars. It has also supported UU Agent, 袁浩’s “cyber digital life” experiment, whose goal is “as a child, you must surpass Claude Code.” In more than 100 days, it evolved from 0 lines to 100K lines of code; the next goal is to “create value for humans.”
• 天杰’s stance toward the next 500 days is to embrace change while retaining “the courage to start over.” Thinking once looked like the main theme, only to be quickly replaced by agents and loops. “Sometimes you hesitate for 1 or 2 months and realize you are already too late.” As for forecasts: “Today’s judgments are all wrong”—looking back 500 days from now, they will certainly be wrong. That is fine.
• 归藏’s closing framework carries the strongest investment implication. To find the next-generation form after AIGC, the logic may need to be reversed: after 10 or even 20 years, front-end app experiences have converged on the forms users find most acceptable. “The product form may not change, while production efficiency rises 100x to 1,000x,” producing entirely different outcomes on the front end. The analogy is pig farming: the shift from household free-range production to industrial farms after electrification. “With capital investment, you can produce higher-quality pork at a lower price, and then major leaders will emerge.” “The next Douyin may be a Douyin with 10x the output.… The next video will still have the form of video; only the way it is produced will be completely different.”