Seal: Xiaohongshu’s path from a 3-day OpenClaw demo to company-wide AI assistants
From a three-day demo to company-wide rollout: Xiaohongshu’s Seal enterprise AI assistant
Meng Hongjin (talk) · edited by Li Zhongliang · InfoQ · 2026-10-29
At AICon, Xiaohongshu architect Meng Hongjin described Seal, an enterprise personal AI assistant. Before Lunar New Year the team forked OpenClaw and shipped an internal beta in three days with three people; after validating demand they spent a week making the architecture horizontally scalable, opened 2,000 public-beta seats that sold out in five minutes, then spent two more weeks scaling to company-wide coverage by mid-March.
They chose secondary development on open source over a greenfield build, citing product maturity, cheap validation, plugin extensibility, and community Skill ecosystems.
Security has three layers: cluster isolation from production with office-network connectivity; per-user Pods with dedicated disks, later a Nex sandbox platform (Kata/microVM, warm pools); sensitive-data detection plus model auto-routing so sensitive files stay on in-house models.
Cost controls include Self-GC (JVM-like context garbage collection; ~10–15% daily token drop, ~20% peak) and Auto routing (~70% of tasks on cheaper models; ~69% model-cost savings vs always using the premium tier).
Memory was prioritized over Skill self-iteration: Session → Worklog fact buffer → Memory Wiki. Skillhub reached 9,000+ Skills within a month (>8,000 internal). CoWork Studio lets non-engineers publish AI-coded tools as shareable internal sites.
Next directions: proactive todo awareness, goal-oriented delivery, and multi-agent collaboration.
https://www.infoq.cn/article/eOqH5GJ5R6UuZGtlhL53Reposted. Will be removed on request. All rights belong to the original author.
More from this batch
Crusoe sells troubled Wyoming data center project to Google
Crusoe sold its unfinished SE Wyoming campus to Google, which plans ~$17B to build it out after cost/timeline concerns.
Measuring how AI agents handle everyday wet-lab science
OpenAI+Ligo BioBasicsBench: 100 raw wet-lab tasks; GPT-6 Astra ~66.7% at xhigh vs 43.8% for GPT-5.6 Sol.
OpenAI×Frontier Model Forum:给「绕过防护是否可能造成严重网络伤害」定级框架
官方帖:与 FMF 成员机构一起发布框架,帮助判断绕过 AI 防护是否可能促成严重网络危害,并优先加强防护。属安全评测方法论,非新模型发布;可与 Vals CyberOffense 等攻防评测对照。
vork by Voyager: a headless 3D and VFX engine for agents
Research-preview headless 3D/VFX engine: agents write scenes as code (~2.5k-token API), render on GPU, inspect, revise.
Claude Duet 手机端可监控与审批 Claude Code 会话(本地与云端)
多条社区与日报称:Claude Duet App 已能查看 Claude Code 会话状态、审批改动、发送更新,覆盖本地与云会话。对「人在回路」远程盯编码 Agent 有直接产品含义;以 Anthropic 客户端实际能力为准。
Linus Torvalds 批强制披露「是否用了 AI」:结果比工具更重要
X 趋势话题:Torvalds 在编码相关讨论中称强制 AI 使用披露「utterly stupid」,强调看结果而非工具。属社区言论摘要,细节以原帖/邮件列表为准;对企业合规「AI 生成标记」政策是反面对照。
建设者把重心放在 Agent Harness(工具/记忆/子代理/校验),而非「完美模型」
趋势摘要:多位建设者强调用 harness 组合让 Agent 能稳定跑数小时任务。可与本期 Harness-S 论文、ECC 等仓库对照;属舆论主题而非单一产品发布。
Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training
Selective sibling-harness rescue transfers complementary behaviors into one policy that still works under the original harness.
VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding
Parallel primitive decoding cuts trajectory error and latency; VFlash-conditioned WAM hits 70.2% on MolmoSpaces suites.
Decoupled Predictive Decision Making:共享条件预测器服务多 Agent 决策
改进什么:多 Agent 各自学预测易冗余、难共享。怎么做:解耦出共享条件预测器,决策头按需使用。HF↑114,适合对照多智能体世界模型/规划栈。
SPECMEM:用跨会话活记忆加速工具调用的投机解码草稿
改进什么:工具调用仍自回归解码,延迟高;现有系统无跨会话草稿库或建完不更新。怎么做:按用户重复行为维护可更新的草稿记忆,服务投机解码。面向部署侧工具 Agent 延迟。
Tools Need Not Speak in Text:工具结果用隐式上下文压缩,少占对话 token
改进什么:工具返回全进文本上下文又贵又吵。怎么做:对工具观测做隐式压缩再注入 Agent。适合长工具链企业 Agent 控成本。