Skill：microsoft agentic harness
POC template for Microsoft Agent Framework agentic harness — skills, MCP, tools system modeled after Claude Code architecture
适用场景
招聘人事
适合角色
介绍
以下内容整理自开源仓库 `MCKRUZ/microsoft-agentic-harness` 的 README / SKILL.md 摘要： # Microsoft Agentic Harness If you've ever used [Claude Code](https://claude.ai/claude-code) and wondered *"how does this thing actually work under the hood?"* — this project is an answer to that question, built on the Microsoft stack. The Agentic Harness is a proof-of-concept that reconstructs the architecture behind modern AI coding agents: the skills system that decides what an agent knows, the tool system that decides what it can do, the context budget that decides how much it can hold in its head at once, the orchestration loop that ties it all together, and the meta-harness that automatically improves the agent's own configuration over time. It runs on .NET 10, uses Clean Architecture, and speaks the same protocols (MCP, A2A) that the broader agent ecosystem is converging on. It's not a chatbot wrapper. It's the plumbing that makes agents feel intelligent. --- ## The Problem Most "agent" implementations are a prompt, an API call, and a prayer. They work for demos. They fall apart when you need an agent that can: - **Use tools safely** — not just call functions, but do so within a security sandbox where file access is restricted to explicitly allowed paths and every tool call flows through content safety middleware. - **Manage its own context** — LLMs have finite context windows, and a naive agent that dumps everything into the prompt will drown before it gets useful. The agent needs to know what to load, when to load it, and when to let go. - **Collaborate with other agents** — not through shared memory or global state, but through real protocols where agents discover each other, negotiate capabilities, and delegate work. - **Be observable** — when an agent makes a bad decision three tool calls deep in a multi-turn conversation, you need traces, not guesswork. This harness solves those problems with real engineering, not abstractions on top of abstractions. --- ## Learn How It Works Six learning resources, depending on what you need: - **[Developer Onboarding Guide](https://mckruz.github.io/microsoft-agentic-harness/)** — Step-by-step walkthrough of the codebase aimed at engineers who are forking this template to build something. 17 pages covering getting running, every config knob, the Clean Architecture layout, a full message-journey trace, the skills/tools/RAG/MCP systems, observability, the evaluation framework, the SkillOpt-port skill-training loop, generative UI, and the Bundle API for calling the harness over HTTP. Read this if you're going to write code.<br /> ↳ **Calling the harness from another system over HTTP?** Go straight to **[Chapter 17 · Bundle API](https://mckruz.github.io/microsoft-agentic-harness/17-bundle-api.html)** — the consumer integration guide, with a hand-written [OpenAPI spec](https://mckru ---
安装方式
1. 打开仓库：https://github.com/MCKRUZ/microsoft-agentic-harness 2. 查找 SKILL.md 或 README 中的安装章节 3. 克隆或下载所需 Skill 子目录到本地 skills 路径（如 `.cursor/skills/` 或 `.claude/skills/`） 4. 重启 / 重载 Agent 客户端 5. 用一条真实任务验证 Skill 是否被正确加载
适用平台
- Claude
- Codex