Hi, I'm Wenjie, an AI application and agent developer. I build AI Agents, Agent Runtimes and AI-powered applications — and I care about the whole loop that makes an agent actually work in production: a solid runtime & harness, well-engineered context, reliable tool calling, and the observability & evals to prove it.
- 🔭 I'm currently building agent runtimes and AI-powered applications
- 🌱 Active contributor to Apache Magpie (Incubating) and 10+ open-source AI projects — 20+ PRs, 6 merged and counting
- 💬 Ask me about agents, tool calling & MCP, context engineering, RAG, Text-to-SQL
- 📫 How to reach me: [email protected]
| Area | What it covers | |
|---|---|---|
| ⚙️ | Agent Runtime / Harness Engineering | The agent loop: planning, execution, sandboxing, approval gates, retries |
| 🧠 | Context Engineering | Context assembly & compression, memory, prompt/state management |
| 🔧 | Tool Calling & MCP | Tool design & integration, MCP servers, function-calling reliability |
| 📊 | Agent Observability & Evaluation | Tracing, eval harnesses, regression testing for agent behavior |
| 🔍 | RAG / Text-to-SQL | Retrieval pipelines and natural-language access to data |
| 🚀 | AI Applications | Shipping full-stack, LLM-powered products end to end |
20+ PRs across the AI agent ecosystem — framework, runtime security, tool calling, memory, and application layers:
| Project | What it is | My contribution |
|---|---|---|
| 🐦 Apache Magpie (Incubating) | AI assistant framework that helps open-source maintainers triage issues, mentor contributors and draft docs | Contributor — 5 PRs (#1390 merged): trim the issue/PR-triage skill families, harden stats against untrusted timestamps, add Maven artifact verification to release checks |
| 🧠 Agno | Full-stack framework for building multi-modal agents | #10600 — reject forged tool continuations and enforce the approval gate on /continue, with regression tests |
| 🌐 browser-use | Lets AI agents drive web browsers | #5919 — support JSON Schema type arrays (e.g. ["string", "null"]) in tool schema conversion |
| 💬 LibreChat | Open-source AI chat app with multi-provider support | #16409 — generate titles for Open Responses API conversations (shared-service refactor + 12 tests) |
| 💾 mem0 | Memory layer for AI agents and apps | #7419 — LLM provider base_url docs master list · #7431 — AWS Bedrock Converse response parsing fix |
📈 More contributions
| Project | Contribution |
|---|---|
| vllm | #58816 — fix the Anthropic messages test for Anthropic SDK 1.x |
| QwenLM/qwen-code | #12697 merged — reject http archive URLs with an actionable error · #12613 — remove dead REPLACE mergeStrategy declarations |
| bytedance/deer-flow | #5774 merged — document runtime environment variables |
| TencentCloud/Octop | #1049 — document backup / mobile / browser-idle env overrides |
| assistant-ui | #7983 — clarify client refs during SSR in store docs |
| laya | 3 merged docs PRs (#325, #313, #247) — AGENTS.md contribution rules, benchmark-table accuracy |
| Project | Description |
|---|---|
| 🤖 xxx Agent Project | TODO — e.g. an agent runtime / harness I designed, with tool sandboxing & approval gates |
| ⚙️ xxx Runtime / Tooling | TODO — e.g. MCP server, eval harness, or context-engineering toolkit |
| 🔍 xxx RAG / Text-to-SQL | TODO — e.g. a retrieval or Text-to-SQL demo with measured answer quality |
⭐ From Wenjie — building agents, one tool call at a time.