AI engineer building reliable agent systems and production-grade developer tools.
10 years in software engineering · Architecture · Evaluation · Reliability
I work where AI prototypes become dependable systems: agent architecture, evaluation, orchestration, and the controls that make long-running autonomous work inspectable and verifiable.
我做可评估、可追踪、可交付的智能体工程，重点是长任务里的 context death 和验收收敛。技术文章与实践记录在 x-nan.com。
longgraph-skill — long-horizon agent skill for Claude Code, Cursor, Codex, and Grok Build. Clean-context supervisor, multi-task ledger loop, verified gates. A markdown library (loop-graph), not a framework.
herdr-agent-quota — live Claude, Codex, Grok, and Agy quota in the Herdr sidebar: 5h + weekly % remaining, reset ETAs, and time-aware quota health. Local-only; never hit a limit mid-task.
| Project | Description | Stars |
|---|---|---|
| longgraph-skill | Long-horizon agent skill for Claude Code / Cursor / Codex / Grok Build — multi-task ledger loop, host-portable, clean-context supervisor, verified gates. Markdown library (loop-graph), not a framework. | 71 |
| herdr-agent-quota | Credential-scoped AI quota, context, and cache in Herdr for Claude, Codex, Grok, Agy, OpenCode, Pi, omp, and Devin. | 66 |
| obsidian-llm-wiki | Compile notes into a linked Obsidian wiki at ingest time with Claude Code, following Karpathy's LLM Wiki idea instead of RAG-at-query. | 10 |
| sherlock-claude | An AI-powered code analysis platform built on Claude Agent SDK. It automatically analyzes codebases, diagnoses errors from logs, and generates targeted fix recommendations — all driven by YAML-configured agents and a reusable plugin skill system. | 6 |
| dsh-plugin-longgraph | DeepSeek Harness community plugin: longgraph / loop-graph / loop-converge authoring skills on ctx.skills | 5 |
| agent-ding | Ding when your coding agent finishes — modular notifications, Zellij layouts, shell helpers | 3 |
| sherlock-openai | A lightweight platform for building and testing OpenAI Agents workflows with FastAPI backend, LiteLLM model routing, and React frontend | - |
- Long-horizon agents — survive context death, keep state durable, make “done” checkable
- Loop converge — bind each loop to a real acceptance gate, not empty code churn
- Quota-aware workspaces — live 5h + weekly remaining in the Herdr sidebar so long tasks don’t die mid-run
- Skill tooling — turn docs and good sessions into installable skills
- Evaluation & reliability — evidence-based acceptance and failure analysis
Make the state inspectable. Make “done” verifiable. Keep the system smaller than the problem it solves.