复盘 2C 和 2B Agent 发展脉络｜对谈 Afk.surf 创始人嘉琛

Summary

This 42章经 conversation with 何嘉琛 traces his path through AFFiNE and AFK, then contrasts consumer personal agents with enterprise agents. Its central thesis is that an enterprise agent should learn across employees, software, and workflows while retaining person-specific memory and execution boundaries; repeated successful work may then be cached or distilled into cheaper task models instead of being rebuilt independently by every employee.

Key Claims

• AFFiNE correctly anticipated a multimodal workspace, but He argues that fixed editors and canvases become constraints as agents gain autonomy across chat, collaboration software, phone, browser, CLI, MCP, and operating-system environments.
• AFK is presented as one highly concurrent company-wide agent with individual memory and runtime separation, not a collection of isolated personal agents.
• Organizational Agent Learning / 组织级智能体学习 can reduce duplicate context loading, spread effective work methods, and turn distributed employee-Agent activity into shared operational learning.
• Model-Distilled Task Memory / 模型蒸馏式任务记忆 separates expensive first-time reasoning from cheap repeated execution: frontier-model trajectories, tool use, results, and preferences can train a routing model that reproduces stable work.
• Enterprise adoption may be easier to monetize because management can connect permissions, token spend, labor cost, output, and business results, although the episode supplies no audited customer ROI or deployment data.
• AI-Native Role Convergence / AI 原生岗位融合 is illustrated by designers modifying real code and product managers specifying functional logic, while a shared codebase and live artifacts replace long handoffs among isolated role tools.
• The source expects fixed workflows, skills, prompts, and traditional software to lose value as models improve, but this is a strong forecast rather than established market evidence.
• Consumer personal agents face privacy, platform access, supply-chain, subsidy, and incumbent-retaliation problems when they attempt to become the user’s primary transaction interface.

Key Quotes

“一百个人读一百遍同一份代码” — the duplication problem used to contrast isolated agents with shared organizational learning.

“肌肉记忆” — He Jiachen’s metaphor for a small model reproducing a previously reasoned task path.

Connections

• 何嘉琛, AFK, and AFFiNE - founder, current company, and prior workspace product grounding the episode.
• Organizational Agent Learning / 组织级智能体学习, Enterprise Agent Memory, and Enterprise Operational Memory - organization-wide learning, deployed memory, and underlying company-context layers.
• Model-Distilled Task Memory / 模型蒸馏式任务记忆, Model Routing Cost Control, and Agent Self-Evolution - trajectory distillation, lower-cost recurrence, and feedback-loop implications.
• AI-Native Role Convergence / AI 原生岗位融合, AI Organization Design, and AI Coding Verification - role redesign, shared artifacts, and the continuing need to verify code-backed delivery.
• Agent Native Software, On-Demand Apps, and Generated Work Interfaces - forecast that software shifts from fixed interfaces toward dynamically assembled task surfaces.
• Platform-Agent Access Conflict, Agent Permission Boundaries, and Enterprise AI ROI Audit - consumer access conflict, enterprise authority, and value-measurement boundaries.

Contradictions

• No settled contradiction is adopted. The source creates productive tension with governed-workflow and enterprise-software pages: it expects models to absorb workflows and weaken software moats, while existing evidence emphasizes auditability, permissions, deterministic controls, process rules, and systems of record.
• AFK’s architecture, model sizes, task-reproduction quality, cache savings, customer outcomes, security posture, and end-to-end ROI are founder descriptions without independent evidence in the supplied structured summary.
• Predictions about abundant AGI-level models, traditional software decline, platform conflict, role convergence, and a near-term Agent commercialization threshold remain source-scoped forecasts.