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

🚥 This week on Crossing, our guests are the team from Paperboy. John Yang, 21, is CEO. Jett Chen, 19, is a freshman at CMU and a founding engineer. The Paperboy team has 12 people, 10 of them engineers, and has raised $4.7 million.

John believes the best way for humans and AI agents to work together probably hasn't been invented yet. While we already have Claude Code, Codex, Manus, and OpenClaw, they're all fundamentally session-based and prompt-based. You open a window, type a prompt, wait for it to finish, close it. Next time, you start from zero.

Paperboy is trying to find a more natural, continuous, and collaborative agent interface and memory structure — agents should learn by watching how you use your computer, organize conversations through IM rather than sessions, and proactively come to you instead of waiting for your prompt.

If you're building AI products, AI infrastructure, or thinking about how agents enter team workflows, we hope this episode gives you something to think about.

---

🎬 Our video podcast is now live on Koji Yang Yuancheng's WeChat Channels, Xiaohongshu, Bilibili, and YouTube.

📒 The transcript is published on the Crossing WeChat official account.

🟢 00:00 Rapid Fire

Age, alma mater, MBTI and zodiac sign, one-sentence intro to Paperboy, funding status, revenue and profit, team size, pre-founder experience

🟢 01:59 The Starting Point: Today's AI Products Don't Work for Me

All of today's AI products share a common problem: you have to manually throw files, emails, and personal information into a dialog box, and then the conversation just disappears.

Three core pain points: collaboration shouldn't work this way, history shouldn't be preserved this way, and proactivity shouldn't work this way.

"The best way for humans and AI to work together probably hasn't been invented yet."

The existing chatbot and agent product forms are the "default answer" of the AI era, and default answers are almost never optimal.

🟢 06:21 Under the Assault of Claude Code

Three opportunities in the agent track: first, letting agents truly learn from user environments; second, being personalized enough to be proactive without being intrusive; third, the user experience must be extremely intuitive — "you shouldn't have to learn how to use it like a new tool."

These three still depend on human teams, and nothing new on the market has broken this framework yet.

🟢 08:01 Two Fundamental Problems with Agents

Cursor and Manus are currently the most successful agent forms, but John says they have two fundamental problems — and this directly defines what Paperboy is building.

Problem one: session-based. You have a bunch of workspaces, a bunch of conversations, and every new session is like meeting someone for the first time; context doesn't follow you around.

Problem two: reactive. You have to ask first, then it answers. Agent.md files need active maintenance.

An agent should know your mouse movements, your video and audio, all your computer activity — context should be far longer than any context window.

🟢 14:21 After Screen Data Became Industry Consensus

"Collecting user screen data to build a context layer has, to some degree, become industry consensus."

Codex and Littlebird are doing it, various players are doing it, but what comes next — whether it's predicting the user's next keystroke or predicting what they'll do in the next hour — no one has found the best "recipe" yet.

This area needs massive amounts of engineering and research, "and today, for a company, exploring this territory is still a very good choice."

🟢 16:46 Mini Vivian & Auto John

Mini Vivian is a Paperboy instance the team trained in their internal Slack, which understands everything Vivian has ever said, her judgment and taste, her hiring standards.

It can help Vivian dig up candidates from GitHub, Xiaohongshu, and Twitter.

Talking to Auto John (John's agent double) is sometimes smoother than going directly to John.

"I look forward to the day I can just lie flat and let Auto John become something smarter than me."

🟢 27:36 WeChat Group Chats Inspired the Interface Design

Different roles would have completely different "sidebars," and if everyone got a customized version, it could never become a product.

The turning point came from WeChat — the same group of people can simultaneously exist in multiple groups with different themes, which is the most natural way humans organize information, and it doesn't feel annoying.

🟢 33:36 The Last Interface and Five Speeds

Paperboy's only blog post so far is titled "The Last Interface" — the speed hierarchy of context determines product form.

The "pace layers" theory: fashion, commerce, and infrastructure each change at different rhythms.

Mapped to AI products: tasks within 1 second might best be autocomplete; tasks of several hours use IM; longer time horizons are "a very worthwhile area to explore."

"Five speeds" doesn't mean Paperboy literally only does five — it's a thinking framework: which speed layer of automation you're operating on directly determines what kind of product you should build.

🟢 42:09 Two Types of Engineers, One Book, One Coach

On a 12-person team, John says he hires two completely different kinds of people.

The first: someone like Jett — young, high-IQ, brimming with creativity, who can rapidly prototype a solution for every hard problem. The second: someone with extremely solid domain fundamentals, like an engineer from AWS who worked on Windows kernel development, to handle the underlying infrastructure.

Management experience is almost entirely self-taught: High Output Management, The Hard Thing About Hard Things, Trillion Dollar Coach — plus one hour a week with a CEO coach who came from a former VC executive background.

A coach is so much better than therapy, because: "you can talk about emotions, then immediately talk about what's actually happening in the business."

🟢 48:07 Turning Down Cognition, Vercel, Sentry — Then What

In the Million era, they received acquisition offers from Cognition (Devin), Vercel, and Sentry.

"Joining those companies, in a way, is like being a regular employee — you have to work on someone else's ideas."

Favorite AI product? Jett: Codex — "it's defining how future software engineers work in its most ambitious form," and the core agent is open source.

---

Subscribe to Crossing: 🚦 We follow the industry shifts and new entrepreneurial opportunities brought by the new wave of AI technology.

Crossing is Steve Jobs's metaphor for Apple — standing at the intersection of technology and liberal arts, where great products are born. AI is transforming every industry. We seek out, interview, and bring together a new generation of AI founders and active builders in the AI era. Together with them, we explore and embrace the new changes, the new possibilities.

👦🏻 Host Koji: I founded Crossing and started AI Hacker House, a community space for a new generation of AI founders. I serve as Venture Partner at ZhenFund. I believe technology, especially AI, is the greatest value creation opportunity of our generation. Koji on Jike, Koji's website

👧🏻 Host Ronghui: I co-founded Crossing. I've worked at a dollar-denominated VC and spent five years as a Silicon Valley correspondent, following tech development and business stories. Feel free to chat with me and exchange ideas. Ronghui on Jike