Dedicated to the engineering of Agentic Workflows and the evolution of Recursive Self-Improvement (RSI). I believe the path to AGI lies in simulated, self-evolving environments rather than static world models.
- 🛠️ Agentic Frameworks: Architecting the governance core of
DeepAgents, focusing onpatchToolCallsMiddlewareand high-fidelityvalidatorToolimplementations. - 🧩 Native Plugin Standards: Driving native plugin specifications for Claude Code and OpenClaw (PR #1093), prioritizing native integration over bridge-based proxies.
- 🔌 MCP Protocol Evolution: Refining the
openclaw.plugin.jsonmanifest and implementing high-performancehttp-streamtransport layers for Model Context Protocol. - 🧠 LLM Internals: Implementing a full-featured Transformer Decoder from scratch in TypeScript to explore the mathematical essence of Universal Approximation and Backpropagation.
- Languages: TypeScript (Engine Level), Python, Go.
- AI/ML Theory: Transformer Architecture, BP Calculus, Neural Network Optimization.
- Agent Infrastructure: DeepAgents, LangChain/LangSmith (Private Deployment), OpenClaw.
- Protocols & Runtimes: MCP (Model Context Protocol), Docker, Node.js/Bun.
- RSI > World Models: I prioritize recursive self-learning environments that allow Agents to evolve through simulation.
- Native over Proxy: Real progress is driven by AI Agents supporting native plugin standards to reduce latency and architectural complexity.
- Proofs over Black-boxes: I value mathematical rigor (FFN/BP) over empirical "it just works" solutions.
- 🏢 Working at: ByteDance
- 💬 Ask me about: Agent middleware lifecycles, MCP
http-streamprotocols, or why RSI is the definitive path to AGI. - ⚡ Fun fact: I prefer reading the mathematical proof of the Universal Approximation Theorem over a 100-page user manual.