What building Shippy taught us about building agents
Ai2's Skylight team shares engineering lessons from Shippy, a maritime-domain-awareness agent serving analysts in 70+ countries, and open-sources the OpenClaw agent framework, the Harbor evaluation framework and the Mothership hosting platform.
Marked as a signal by the editors; worth tracking.
On this page11 sections
The editors' notes below — why it matters, the context, the learning path — are written in Chinese.
Record
The signal
#Shippy is an AI agent built by Ai2's Skylight team for real-time maritime domain awareness, assisting analysts across 70+ countries in high-stakes operational decisions. The writeup distills four lessons: separate an agent into soul (system prompt), skills (markdown-documented capabilities) and config; talk to APIs through a purpose-built deterministic CLI rather than raw calls; isolate each user session in its own sandboxed Kubernetes deployment; and evaluate on live data with expert-weighted rubrics instead of static benchmarks. Released artifacts: the OpenClaw agent framework, the Harbor evaluation framework with a Shippy plugin, the Mothership hosting platform, and the agent-skills markdown spec.
Why now
Why it matters
#生产级智能体的公开工程复盘本就稀少，高风险场景（错误决策有真实代价）的更是罕见。四条经验——三分解剖、确定性工具层、会话级沙箱、专家评分的活数据评估——每一条都指向同一个主题：智能体可靠性来自工程结构，不是模型能力。
Background
Technical context
#「模型通过确定性 CLI 而非裸 API 与系统交互」是近期智能体工程的收敛趋势：把不可靠性关进模型与工具的边界之内，工具层用传统软件工程的确定性来兜底。agent-skills 的 markdown 规范与 Claude Code、Codex 通用，也印证了智能体能力描述格式正在跨厂商标准化。
Audience
Who should care
#Next
Learning path
#- 对照 soul/skills/config 三分法审视自己的智能体：提示、能力、配置是否已经解耦
- 把最容易出错的一个 API 调用包成确定性 CLI 工具，对比错误率变化
- 参照 Harbor 的思路，为自己的智能体定义一份领域专家加权的评分表
Network
Where it sits in the network
#How this signal connects to neighbouring technologies, skills and background knowledge. Follow any link to keep exploring.
Further thought
Follow-up questions
#- OpenClaw 与现有主流智能体框架相比的差异化设计是什么？
- 会话级 Kubernetes 沙箱的成本与冷启动延迟在什么量级？
- agent-skills 规范会不会成为跨厂商的事实标准？
Publication record
Featured in
#The briefings that carried this signal, newest first; the last row is the weekly review of its week. The tag is the section it ran in.
技术简报 - 2026-08-04
Worth tracking技术简报 - 2026-07-30
Worth tracking技术简报 - 2026-07-19
Worth trackingWeekly review · Jul 13 – Jul 19
Worth tracking
Sources