人类和 AI Agent 的最佳配合方式，还没被发明｜对谈 Paperboy
Summary
This Shizilukou Crossing episode interviews Paperboy founder Jiang Yang and founding engineer Jie Dechen about why the best form of Human-Agent Collaboration has not yet been invented. The discussion argues that useful agents need OS-Level Context, Persistent Agent Memory, and Proactive Agents rather than isolated chat sessions and repeated prompting. Paperboy’s product exploration moves from “AI Slack” toward IM/inbox-like interaction, meeting prep, OS-wide autocomplete, recruiting support, PR descriptions, and an agent layer that works where users already work.
Key Claims
- Current agent products are too session-based: once the chat or project session ends, valuable context, preferences, and taste are often lost.
- Paperboy’s core bet is that agents should learn from the user’s chosen computer environment: screen activity, keyboard and mouse behavior, meetings, messages, search, browsing, and other OS-level signals.
- Persistent Agent Memory is presented as a continuously updated understanding of the user’s profession, recent work, immediate activity, relationships, and decision heuristics.
- The first useful experiences may be narrow but context-rich: meeting preparation, OS-wide autocomplete, commit messages, PR descriptions, recruiting research, efficiency feedback, and connecting research threads.
- Jiang Yang frames prompting as an unnatural bottleneck. The desired interaction is closer to a high-bandwidth working relationship where the agent already understands enough context to help without repeated explanation.
- Paperboy moved away from replacing Slack with an “AI Slack” because enterprise switching costs, integrations, external connections, and network effects are heavy; the more plausible path is to bring agents into existing workflows.
- The team sees OpenAI and Anthropic as the most important competitive forces because they control models and distribution. Paperboy hopes to compete through product taste, new interface forms, enterprise use cases, and user adaptation.
- The source treats model-company and product-company boundaries as increasingly blurry, using Cursor as an example of a product company that may also train specialized models.
- The startup is early and unproven. The transcript records zero revenue, negative profit, a team of 12 full-time employees, and a 4.7万美元 funding figure, but those figures are captured from the transcript rather than independently verified here.
Key Quotes
“人类和 AI 配合工作的最佳方式，很可能还没有被发明出来” — Paperboy’s internal thesis as cited by the host.
“我讨厌 prompting” — Jiang Yang’s shorthand for why agent interfaces should move beyond explicit prompt management.
“把 agents 带到人们已经在做事的地方” — the adoption strategy behind complementing Slack and OS workflows.
“不同时间尺度” — the product-design frame behind autocomplete, meeting prep, and longer-horizon agent work.
Connections
- Paperboy — central company trying to invent a new human-agent collaboration form.
- Jiang Yang — founder describing the product thesis, interface pivots, team-building, market choice, and competition.
- Jie Dechen — founding engineer explaining OS-level memory, autocomplete, PR descriptions, and AI-tool preferences.
- Millian, Same.Dev, and EarthKit — background projects that explain the founders’ prior work and taste.
- Human-Agent Collaboration, OS-Level Context, Persistent Agent Memory, and Proactive Agents — core new concepts added by this source.
- Agentic Workflow, Context Engineering, Agent-Facing Interfaces, and Headless Software — existing agent concepts extended by the episode.
- Slack and WeChat — communication/workflow contexts where agent interfaces may sit on top of existing behavior.
- OpenAI, Anthropic, Cursor, Codex, Claude Code, and Manus — agent and model-company reference points used to frame the competitive landscape.
- Digital Employees — adjacent enterprise frame: Paperboy’s personalized agents also need onboarding, responsibility boundaries, and permission design.
- AI Organization Design — the source discusses founder leverage, market picking, recruiting, management learning, and the continuing importance of deep human experts.
- Embodied AI and AI Governance And Compliance — adjacent opportunity areas raised through Jie Dechen’s interest in robotics and AI security.
Contradictions
- No direct contradiction with prior wiki content. The source reinforces earlier claims that durable agent value depends on context, workflow integration, and interface design rather than model capability alone.
- Tension with Headless Software is complementary rather than contradictory: Paperboy is not arguing GUI disappears, but that human-facing surfaces may become IM, inbox, autocomplete, or review layers on top of agent behavior.