Liangcai Li
AI Application Architect at Youdao, core R&D contributor to LobsterAI.
A Technical Expert in the Intelligent Hardware Division at Youdao and a core R&D contributor to LobsterAI. He has been deeply involved in the cloud service architecture of multiple hardware products, including Youdao Dictionary Pen, Q&A Pen, and Listening Assistant. He has successfully delivered backend capabilities for several AI-powered education products, such as “Little P Teacher” and video-based Q&A systems. Currently, he leads the full-stack architecture design for LobsterAI, driving its evolution from a vertical education-focused agent into a general-purpose, all-scenario agent. His work spans the complete technology stack, including Electron desktop applications, web applications, and cloud services. He has also played a key role in the technical selection and implementation of migrating from the Claude Agent SDK to the OpenClaw engine. He possesses extensive experience in high-concurrency service architecture, AI agent system design, multi-agent collaboration, and the engineering practices of Vibe Coding.
Topic
The Development and Practice of LobsterAI
LobsterAI, developed by NetEase Youdao, is the first 100% open-source, all-scenario personal assistant agent released by a major tech company in China. Built on the OpenClaw engine, it deeply supports everyday productivity scenarios such as data analysis, document processing, PPT generation, video creation, and email workflows. In this talk, we will share the full journey of how LobsterAI evolved from a vertical education-focused agent into a general-purpose, all-scenario agent. We will provide an in-depth look at the architectural transition from a self-developed system based on the Claude Agent SDK to the OpenClaw engine, including key technical decisions and real-world implementation practices. Topics will cover startup optimization, MCP Bridge design, multi-agent architecture, plugin development, as well as engineering insights and lessons learned from adopting the Vibe Coding paradigm. Outline 1. LobsterAI Product Overview — Core Capabilities of an All-Scenario Agent One-command execution, Skills & MCP, scheduled tasks, multi-agent collaboration, IM-based remote control, memory system, and real-world use cases. 2. The Origin of LobsterAI — From Vertical to General-Purpose Agent The evolution from education-focused agents such as “Little P Teacher” and video Q&A systems to a general-purpose assistant, driven by real user needs beyond technical audiences. 3. Architecture Choices and Technical Practices Strengths and limitations of the self-built architecture based on Claude Agent SDK Migration to the OpenClaw engine: decision-making and implementation (independent process + patch mode) Windows startup performance optimization (ESM single-file bundling, ChildProcess replacement, V8 compile cache) MCP Bridge design and implementation OpenClaw plugin development (AskUserQuestion authorization interception) Layered design of multi-agent architecture 4. Vibe Coding in Practice The “sweet spot” and challenges of AI-assisted coding; engineering practices including SDD (spec-driven development), regression testing, and design reviews. 5. Future Roadmap From chatbot to Agent OS; Skills-as-a-Service, open ecosystem collaboration, and new opportunities for AI-native workflows.
