Glassbox Agent Harness
Glassbox-Agent-Harness面向AI Agent 的飞行记录器与评测实验室,让每一次行动可观察,每一份上下文可追溯,每一次优化都有证据。
Selected Work · 精选作品
这里收录可直接查看的 Agent 基础设施、开发者工具、产品原型与开源实验。
每个案例都提供产品说明、仓库、站点或贡献记录;完整同步目录仍可在 Archive 中按需查阅。
01 — 012
面向AI Agent 的飞行记录器与评测实验室,让每一次行动可观察,每一份上下文可追溯,每一次优化都有证据。
从这里开始:查看产品能力、运行记录、仓库与关联笔记。
Agent Arena — 让 Agent 在真实任务中比赛、被裁判审查、留下证据、积累声誉的智能体竞技场。Real-task evaluation harness for LLM agents with judge review and reputation.
AI-native software engineering harness — 18 agents, 9 workflows, one closed loop: Issue → Worktree → Plan → Build → Review → Evidence → Merge → Memory.
A low-token visual evidence compiler for text-only coding agents. Convert images into compact Visual Evidence Packets (VEP) for DeepSeek, Codex, Claude Code, and other text-only models.
Autonomous multi-agent economic sandbox — A2A protocol agents compete, trade, and govern inside a simulated economy with a God's Eye observability dashboard.
Personal second-brain dashboard — Obsidian vault sync, knowledge graph, daily notes, and task tracking in one local-first UI.
多模态人体状态感知 → 中医体质分析 → 智能按摩方案推荐 | OPLRI + TCM + EEG pretrained models · Gradio demo
AMD Hackathon contribution: an AI-native live-event poster engine that combines music rhythm, stage lighting, and real-time visual composition.
An interactive Chinese source-code guide to LLM inference: 13 browser experiments make continuous batching, KV cache, paged attention and scheduling visible.
A personal trading operations dashboard that brings market data, positions, backtests and execution checks into one TypeScript workspace.
An image-to-collectible skill that turns a supplied subject into a faithful, commercially believable packaged object.
On-chain agent intent verification for a Monad hackathon: deterministic evidence checks followed by a fail-closed human signing gate.
Content Graph
A curated trail from shipped projects to the writing and channels that explain how they were made.
Content Directory
Browse the writing, channels, projects, and selected external contributions that currently have a clear path into the public story.
GitHub sync Aug 26, 2026 · 4/7 allowlisted contributions · no warnings
Showing all 21 entries
No matching content entries.
Theme Roadmap
A curated view of the technical themes connecting projects, writing, and the channels where the process is shared.
可观察、可评测、可迭代的 Agent 运行时与工程基础设施。
视觉证据、图片生成与视频生成能力的低成本实践。
从产品设计、系统架构到可运行体验的全栈产品构建。
面向开发者的工具、教程与快速验证型实验。
Find a build · 找一个构建
共收录 64 个项目
Featured / 01
面向AI Agent 的飞行记录器与评测实验室,让每一次行动可观察,每一份上下文可追溯,每一次优化都有证据。
Selected work
Agent Arena — 让 Agent 在真实任务中比赛、被裁判审查、留下证据、积累声誉的智能体竞技场。Real-task evaluation harness for LLM agents with judge review and reputation.
AI-native software engineering harness — 18 agents, 9 workflows, one closed loop: Issue → Worktree → Plan → Build → Review → Evidence → Merge → Memory.
A low-token visual evidence compiler for text-only coding agents. Convert images into compact Visual Evidence Packets (VEP) for DeepSeek, Codex, Claude Code, and other text-only models.
Autonomous multi-agent economic sandbox — A2A protocol agents compete, trade, and govern inside a simulated economy with a God's Eye observability dashboard.
Personal second-brain dashboard — Obsidian vault sync, knowledge graph, daily notes, and task tracking in one local-first UI.
多模态人体状态感知 → 中医体质分析 → 智能按摩方案推荐 | OPLRI + TCM + EEG pretrained models · Gradio demo
AMD Hackathon contribution: an AI-native live-event poster engine that combines music rhythm, stage lighting, and real-time visual composition.
An interactive Chinese source-code guide to LLM inference: 13 browser experiments make continuous batching, KV cache, paged attention and scheduling visible.
A personal trading operations dashboard that brings market data, positions, backtests and execution checks into one TypeScript workspace.
An image-to-collectible skill that turns a supplied subject into a faithful, commercially believable packaged object.
On-chain agent intent verification for a Monad hackathon: deterministic evidence checks followed by a fail-closed human signing gate.
Recent & maintained
Personal portfolio site built with Svelte/SvelteKit — dark mode, motion-rich sections, deployed via GitHub Pages.
一个可安装的个人 Agent Skills 集合,覆盖个人站发布、AI 工程、内容表达与开源工作流,可通过 npx skills add 安装到 Codex 等兼容 Agent。
Living Answer helps readers see which important premises in an older Zhihu answer have materially changed, what that means today, and which evidence supports the update.
https://lora-sys.github.io/nano-sglang-interactive-guide/
A reusable Agent Skill for producing polished standalone teaching HTML pages with content-specific SVG diagrams and advanced but restrained motion.
Shape Up Project Shaping — 把模糊想法塑形成值得下注、边界清楚、可验证的项目。基于 Basecamp Shape Up 方法论,适配 AI Agent 时代的项目塑形 Skill。
面向成长型 B2B SaaS 的可信证据控制层,将政策、系统控制、历史承诺、风险审查与人工责任连接成可审计的承诺链。
MonadMon — the first living creatures on Monad. On-chain creature-raising + PvP battle. Tamagotchi meets crypto.
StepFun image generation backend for Hermes Agent — step-image-edit-2, step-2x-large, step-1x-medium with pluggable API.
MiniMax (海螺) image + video generation backends for Hermes Agent — image-01, Hailuo-2.3/02/S2V-01.
Global system cleanup and management TUI for Linux — services, Docker, Flatpak, disk usage, startup audit, all from one terminal.
Skills management dashboard for AI workflows — browse, install, version, and audit Claude/agent skills across projects.
Complete public record
UI aesthetic improvement workflow — opinionated Claude skill for taking functional UIs to portfolio-grade visual quality.
AI Company Operating System — concept for an autonomous org where agents own P&L, ship product, and serve users end-to-end.
Init-codebase Claude skill — bootstraps any repo into a Claude-Code-ready workspace with rules, agents, and skills wired up.
My NeetCode.io algorithm problem submissions — Python solutions with explanations, complexity, and test runs.
ETH Beijing hackathon entry — original submission archive with code, demo, and write-up.
Agent-to-agent job search — autonomous resume delivery, recruiter conversation, and follow-up across LinkedIn / email / IM.
Distilled Claude skills for Marvis — my personal toolbox of focused agent capabilities.
Custom Your Agent!!
LaTeX + KaTeX-based resume generator
RecruitDraft — 输入公司文化 + 职位描述,AI 生成 STAR 法则面试问题和 LinkedIn 职位帖。A recruiter's copilot for STAR-rule interview kits and LinkedIn job posts.
Science-paper writing platform
Microservice practice project — service decomposition, gRPC + REST gateways, observability, and deployment.
Redis client/server implementation with RESP protocol and data persistence
科研全流程 Claude skill — 文献检索、学术脉络梳理、论文写作润色、技术路线图生成。Full research workflow assistant.
Humanize-write Claude skill — strips AI tells from generated text, returning natural, voice-preserving prose.
YouTube-style video clone with AI recommendations — full-stack Next.js + recommendation pipeline.
node-base — n8n-style AI workflow builder clone with node-graph editor and execution engine.
This project is for learning the nanochat concept
zchat-demo — modern chat UI demo built with Next.js, streaming responses, multi-model support.
Hackathon-tuned AI agent helper — optimized for time-boxed builds: scaffold fast, ship features, win.
Mianshihou (面试猴) — interview prep website backend with Node.js, question bank, mock interview, and progress tracking.
Personal website — early version of lora-sys portfolio.
Meme Vibe Casino — vibe-coded casino game for hackathon, meme-token themed slots and mini-games.
Monad hackathon demo project — quick prototype exploring Monad's parallel execution.
Web3 frontend UI/UX reviewer Claude skill — audits dApp interfaces for wallet flow, signing UX, and on-chain clarity.
ui-design Claude skill — generates component systems, color tokens, and typography scales from brief.
Better-auth library learning project — auth patterns, session management, OAuth flows.
Go learning sandbox — small projects exploring goroutines, channels, interfaces, and std lib.
Mianshiya-next — interview practice platform clone (Java backend).
Daily RSS — minimal daily RSS reader with smart summarization and topic clustering.
Interactive Sci-Fi Control Interface built with Three.js — WebGL experiments, shader work, scene composition.
BI project — dashboards, metrics pipelines, and visualization experiments.
Generic demo project — template for prototypes.
Note-app — personal note-taking app with markdown, tags, and quick capture.
Dream-prop random generator — spin up random D&D-style props, characters, and scenarios.
React practice — components, hooks, state patterns, and small apps.
C++ practice projects — algorithms, systems, and small utilities.
Automatic Compliments (自动夸人小助手) — generates personalized compliments in Chinese and English.
React projects — collection of React apps and components.
Snake — a very interesting snake game in C++.
换个关键词、分类或状态试试。