The Best Way for Humans and AI Agents to Work Together Hasn't Been Invented Yet | A Conversation with Paperboy

Summary

• Paperboy’s core bet is to build agent memory by observing users at the OS level. That means screenshots, keyboard activity, mouth movements, meeting audio and video, browsing history, and iMessage, with access granted by the user. Jet’s comparison is straightforward: “Watching someone use a computer for roughly 60 minutes can teach you a lot about them,” while looking only at 60 minutes of WeChat messages “doesn’t actually teach you much.” The team initially tried exporting complete WeChat and iMessage histories, quickly found that approach unscalable, and moved up to OS-level data collection.
• Screen capture is now an industry consensus; the edge is in the pipeline. 酱阳 calls it “the most obvious next frontier for context”—Codex, Clockwork, and Clock Code will all do it. The host said OpenAI has probably also released Chronicle. How raw activity streams are compressed and structured depends on the application: the “action space is large enough,” and for proactive agents—whether they predict the next keystroke or what the user will do over the next hour—“nobody has found the absolute best recipe” yet.
• The product starts with a critique of today’s dominant agent paradigm. Claude Code and Manus are the most successful agent forms, but both are session-based and prompt-based: valuable insight embedded in one of thousands of sessions disappears unless deliberately saved, while maintaining .md files “makes it very hard to translate all your taste and judgment into raw text files.” 酱阳 puts it bluntly: “I hate prompting… people don’t think in prompts.” Paperboy replaces sessions with a persistent, searchable, iMessage-like chat stream.
• The competitive read is blunt: only OpenAI and Anthropic matter today. Everyone else is behind. The only path for a startup is to build a new interface early and have a Cursor-like edge in taste. The AI Slack category faces two structural problems: switching costs for enterprises and agent learning rates. An agent that learns only from messages sent to it will not learn enough, and “people with jobs are busy—they don’t have time to train agents.”
• 酱阳’s public-market allocation is OpenAI 80% and Anthropic 20%. OpenAI’s infrastructure advantage lets it subsidize more compute, and its values are more unopinionated. Anthropic is compute-constrained and “doing a lot of weird things”—a repo containing open cloud or Hermes agent could potentially be charged 10x the price. 酱阳 would buy neither at current valuations. On tools, he prefers Codex to Claude Code: “The most important thing about Claude Code is showing how good Opus is, not how good its command line is.”
• With a hypothetical $3M to invest, 酱阳 names Slack, robotics, and infrastructure. Another speaker, not clearly identified in the source, also picks robotics and adds AI-plus-security: demand is “basically infinite,” offensive and defensive capabilities improve with the models, and the market is extremely meritocratic—uncovering 10% more vulnerabilities can make a product win. 酱阳’s call is categorical: the next company on Cursor’s scale will come from consumer, and “it’s not going to come out of enterprise AI.”
• The company has raised $4.7M and is “at 25,” though the original does not specify what 25 refers to. It has “zero and negative gross margins—we lose money every day,” 12 full-time employees, and 10 engineers. It previously rejected talent-acquisition approaches from Cognition and from the entities heard in the original subtitles as Brazil and Century, whose identities are unclear: “If you can stay independent and achieve your dream, you stay independent. That’s what every founder wants.”

Deep dive

1. The Opening Thesis: The Best Collaboration Model Hasn’t Been Invented—and It’s a “Moving Goalpost”

• 21-year-old 酱阳—possibly at Pratt Institute studying architecture, a 2x founder whose first company was Million—wrote the thesis in the company’s first all-hands document: “The best way to work with AI hasn’t been invented, and we have a shot at being the team that figures out that way.” At the time, on November 30, the company had only 4 people. The reference point was Cursor, the first company to truly lock onto—and broadly succeed at—how to code with AI.
• The past 6 months brought one shock after another—Cloud Code, Open Cloud, Arms—and left many founders “feeling desperate.” But 酱阳 says the direction has “not really changed dramatically,” because the problem space comes from people themselves, and the goalpost keeps moving: every time you come up with something new, everyone else sees it, and all you have to do is keep getting better.
• He breaks the problem into 3 requirements. Agents must learn autonomously from the user’s environment, where data lives in computer files and integrations. They must be personalized, requiring fewer prompts, capable of being entrusted with higher-stakes decisions, and reliable over long periods. And the design must be intuitive enough to require no learning, with a form factor capable of carrying proactive output.

2. The Critique of Session-Based Agents—and What Paperboy Is

• 酱阳’s diagnosis of Claude Code, Manus, and similar agents is that session-based systems leave workspaces full of thousands of sessions: “Some sessions contain extremely valuable insight in their context window, but if I don’t deliberately save it, that session is gone.” Prompt-based systems are reactive and try to compensate through skills and .md files, but “no matter how much effort you put into maintaining them, it’s very hard to translate everything about your taste and judgment into these raw text files.”
• Paperboy responds in 2 ways. First, the agent learns by watching the user work on a computer—screenshots, keyboard activity, mouth movements, meeting audio and video, browsing history, and iMessage, with the user choosing which permissions to grant. Second, conversations live in persistent, searchable chats that extend beyond the context window. The product feels like opening iMessage or WeChat: a list of chats, with each conversation available to resume.
• The hook in the first hour is meeting preparation. Once the user connects a calendar and email, the agent reads along while asking who people are, then proactively suggests: “I see you have this meeting coming up. Want me to research it?” 酱阳 credits Interaction, the company behind Poke, with having “figured out the recipe”: messaging users proactively without becoming annoying, and establishing the expectation that the agent will text you when it makes sense.

3. The OS-Level Bet: 60 Minutes of Computer Use Beat 60 Minutes of Chat History

• Jet, the 19-year-old founding engineer, CMU freshman, and CTF player, retraces the product’s path. The earliest prototype tried to export a user’s entire WeChat and iMessage history, but the team “quickly realized this wasn’t a scalable way” to proceed. The scalable layer was the OS, based on an information-density thesis: “Watching someone use a computer for 60 minutes can teach you a lot; looking only at the WeChat data from the last 60 minutes actually doesn’t teach you much.”
• The technical foundation is a real-time markdown document. It records the user’s profession, what they did over the past few days, and what they did over the past few seconds or minutes. The closer the data is to the present, the finer the granularity.
• The candid status update is that the first prototype shipped to friends had cost problems: “It’s too expensive. It doesn’t really work.” The next iteration is due to wrap this month, followed by another go-to-market attempt.

4. Screen Data Is Becoming Consensus; The Difference Is the Pipeline and the Application

• 酱阳’s view is that capturing raw data from a Mac and turning it into memories is “the most obvious next frontier for context.” This is not limited to startups such as Air Jelly: Codex, Clockwork, and Clock Code will all do it, and the host said OpenAI has probably released Chronicle as well. But how the raw stream is processed is directly tied to the agent’s application. A company focused on learning how users reply to email will build a different memory structure from one focused on another workflow. The action space is broad enough that there is still room for someone to be first.
• Jet adds that this is a barely explored domain. To build the best proactive agent, are you predicting the user’s next keystroke, or what they will do over the next hour? “Nobody has found an absolutely best recipe” yet, which leaves startups room to explore.

5. The Team’s Aha Moments: From AutoJohn to “I Haven’t Used Claude Since February”

• Every employee has an agent counterpart living in Slack. 酱阳’s is AutoJohn, which answers product and design teams around the clock. Vivian, who runs recruiting, has MiniVivian: “It understands everything I’ve said in meetings and Slack, including my judgment and taste around what kind of candidates we should hire.” It can proactively surface candidates from GitHub, Xiaohongshu, and Twitter. The result: “Vivian hasn’t used Claude since February—you can’t use Claude to research candidates for you.”
• 酱阳’s anti-prompting manifesto is worth preserving: “I hate prompting… people don’t think in prompts; we send text and expect the other person to know what we’re talking about.” He pairs it with an unusually candid ambition: “Today’s models are smarter than we are.” His goal is eventually to lie back and let AutoJohn become an entity with a higher IQ than his own.
• Jet’s reservations remain. Chatting with AutoJohn is “often better than chatting with John,” but there is a concern that its intent does not yet fully represent 酱阳’s and that it may misunderstand. 酱阳’s answer is that users will ultimately be responsible for their agents. During onboarding, the agent asks, “How much can I share with this person?” It defaults to mirroring the user’s sharing habits: 酱阳 shares a lot with Jet and less with a new engineer, and the agent learns that distinction by observing his chats.

6. @pb Global Completion—and the Design Conversation That Turned IM Itself Into an Inbox

• Jet’s earliest aha was OS-level text completion. In any input field, type the activation word @pb, with a short instruction or nothing at all, and it guesses what you want it to do—like a colleague who understands from a glance. In a terminal, @pb commit writes the entire commit message. When opening a PR, it combines cross-application context—Claude Code sessions, WeChat discussions with John, and browser research—into a draft that is “much better than any agent based on a single application’s context.”
• There is also a self-deprecating use case. Jet suspects he has ADHD: he opens a talk, starts browsing Hacker News, and loses 30 minutes. Every night he asks Paperboy for a productivity report, “basically to scold me about when I’m not being productive.” Knowing that it is watching him work makes him more efficient.
• 酱阳’s design case started with VCs as the ICP. They need a personal CRM, a meetings module, quarterly deals, and portfolio tracking. But operators, founders, and real-estate salespeople would each need a different sidebar. Shipping skills, plugins, and recipes and asking users to choose among them “didn’t make a lot of sense to me.”
• Paperboy brought him back to IM. What application has an infinite list without becoming annoying? WeChat. 3 people can open 4 groups for different topics, and inactive groups naturally sink. “My WeChat is sort of like this inbox.” The conclusion became the philosophy of the latest interface: “We don’t design an inbox feature; we turn the thing that is itself an inbox—IM—into the product.” The host offered a live example: Crossroads creates a new group for every podcast episode.

7. The Competitive Landscape: Only 2 Matter, and AI Slack Has a Structural Weakness

• 酱阳’s view of the competitive landscape is unequivocal: “Only 2 that matter today are OpenAI and Anthropic. Everybody else is behind them.” They have the models, new compute, and distribution. Startups can only hope to find a Cursor-like edge in taste and ship a new interface first. The individual-product-experience layer is not going away.
• His critique of the AI Slack category, whose Richard and team he says he respects, has 2 parts. First, how does an enterprise switch? It would need to export and import Slack data 1:1 and rebuild every connector. Second, the agent’s learning rate: “If the only source from which an agent learns how to work with people is the messages you send it,” that is not enough. “People with jobs are busy. They do not have time to care about training agents. They expect things to work out of the box.” That is why Paperboy abandoned its original AI Slack concept.
• Jet adds Slack’s moat. External Connections mean communication among serious companies has already moved around email and onto Slack. That WeChat-like network effect is hard to replace. An agent operating at the OS level and complementing Slack “actually has a better adoption story.” 酱阳 confirms the strategy: bring the agent to where people already work instead of rebuilding everything from zero.

8. The Last Interface: Pace Layers and the Spectrum of Task Duration

• The ideas behind the 5 kinds of speed in the company blog The Last Interface come from Long Now Foundation founder Stuart Brand’s Pace Layers: fashion, commerce, politics/governance, infrastructure, civilization, and nature—6 layers that change at different speeds and interact to form society. 酱阳 has been fascinated by the framework since high school: in the agent world, “you also need these different pace layers.” He emphasizes that “5” is simply a sensible classification, not a claim that there are only 5 layers.
• In product terms, this becomes a spectrum of task duration. At one end is replying to a WeChat message in 1–10 seconds, where auto-complete is already operating. At the other is spending several hours reading 10 long reports to make a business decision, and beyond that are tasks measured in months. 酱阳’s view is that every end can be augmented or automated by some kind of agent, but the job and product form will differ. The longer the time horizon, the less certain the form factor: automating a task that takes several hours remains a wide-open area of exploration.

9. A 21-Year-Old First-Time CEO: Management Lessons, Acquisition Offers, and a Rational Decision to Stay in College

• 酱阳’s first CEO lesson is blunt: “Man, it’s fucking hard. I got to own everything.” He is also “the person who can make the company lose the most money, because I have the most leverage.” His solution is to make defining success and ensuring that everyone’s context is complete and accurate his top job. Methodologically, he sees market selection as the most important early founder trait. The market must be large and capable of expanding for 10 years. Historically, the most successful products were never single products; they were product lineups.
• He learned management on the job. At a large company whose name sounds like “manis” in the original audio, he held one-on-ones with CTO Pan Pan and asked every question he could think of. His reading list includes Andy Grove’s High Output Management, which he calls “the best management book ever,” Ben Horowitz, and Bill Campbell’s Trillion Dollar Coach. He also hired a CEO coach for 1 hour a week—a former VC executive who specializes in first-time CEOs. “A coach is much better than a therapist: you can talk about emotions and business,” and Paperboy listens to the coaching calls and helps him follow up on commitments.
• His acquisition history dates back to Million, which received approaches from Cognition and from entities heard in the original subtitles as Brazil and Century, whose identities are unclear. He rejected them because these were talent acquisitions: the money would not be large, and he would become another employee working on someone else’s idea. “If you can achieve your dreams by staying independent, you stay independent.” His hiring philosophy is equally sharp: no hard sell and no scripts. “You can’t talk to mediocre people.” What matters most is the decisions candidates have made in their lives, because “you can’t fake that.”
• One footnote to the anti-Silicon Valley narrative: Palantir is telling high-school students not to attend college, so why is Jet still at CMU? 酱阳 answers first: “It’s got Chinese parents.” Jet’s serious answer is that for most people uncertain about the future, college is a good 4-year option for exploration. But if you already know what you want to do and the opportunity is there, dropping out can be a rational choice. As for following 酱阳, he uses a founder analogy: you have an intercept and a slope; the slope matters much more than the intercept, and 酱阳 is a high-slope founder.

10. Investment Lightning Round: Codex Over Claude Code, OpenAI 80/20, and the Next Cursor Coming From Consumer

• 酱阳’s case for Codex is comprehensive. The core agent and agent loop are open source on GitHub. The desktop product is exceptionally well crafted: OpenAI acquired the Sky team that built Apple Shortcuts, and Codex Pets and computer use are “extremely polished.” Codex was also the first to make highly parallel work the primary UI. Finally, “OpenAI has a much bigger advantage than Anthropic at the infrastructure layer—the models are roughly comparable, but OpenAI is much more stable.” His one-line verdict on Claude Code: “Its most important function is showing how good Opus is, not how good the command line is.”
• On the public-market allocation question, 酱阳 chooses OpenAI 80% and Anthropic 20%. He appreciates Anthropic’s interpretability and societal-impacts teams, but finds the company’s views on how users should use models too opinionated. As a self-described libertarian, he prefers OpenAI’s unopinionated stance. Anthropic’s compute constraints are also pushing it into “a lot of weird operations”—for example, a repo containing open cloud or Hermes agent could potentially be charged 10x the price. He would buy neither at current valuations.
• On the idea that models will swallow everything, his response is that many companies will offer roughly comparable models, leaving product differentiation with product companies. “That may not necessarily be true, but it’s possible.” Cursor is the counterexample: it is both a product company and a model company, and 酱阳 says it could use xAI’s compute to build a frontier coding model. “Is a model company still a product company? It’s usually both.”
• Asked to invest a hypothetical $3M across 3 teams, 酱阳 picks Slack, robotics, and infrastructure, and stakes out the most tradable view: “The biggest company similar to Cursor is going to come out of consumer, not enterprise AI.” Another speaker, not clearly identified in the source, also picks robotics and adds AI security. As models improve, attackers and defenders both get stronger, so demand is “basically infinite.” It is a “highly meritocratic market”: if your product finds 10% more vulnerabilities than competitors, it is highly useful. Geopolitically, countries with the strongest models have an incentive to monopolize the capability, while other countries still need these products. A year from now, 酱阳 wants 2 things from Paperboy: to stop losing money every day and to have higher talent density than today.

Verification Notes

• The raw material does not clearly connect “Melanie was in YC Winter 2024” to Million, so the YC affiliation was not retained.
• The raw material does not explain what “at 25” refers to.
• The entities heard as Brazil and Century, as well as the entity heard as “manis,” cannot be identified from the raw material.
• The raw material does not clearly identify the speaker who added the robotics and AI-security views, so those comments were not attributed to Jet.