Process Engineering for AI Coding

让 AI Agent
遵循你的
工程流程 Make AI Agents
Follow Your
Engineering Process

不要只写更好的 Prompt。开始构建可复用的工程流程。Agent Harness 让 Claude Code、Codex、Cursor 等 AI 编程工具，在不同项目中遵循统一的开发方式。 Stop writing better prompts. Start building better processes. Agent Harness makes Claude Code, Codex, Cursor and other AI coding agents follow a repeatable engineering workflow.

它不是另一个 Agent，而是包裹在 Agent 外面的工程流程层。 Not another agent. A process layer around the agents you already use.

快速开始Get Started → 为什么需要它Why it matters
无需服务器No server只是一组仓库文件Just files in your repo
无需 SDKNo SDKAgent 能读文档即可Works with doc-reading agents
无厂商绑定No lock-in使用你已有的工具Bring your own agent
使用你熟悉的 AI 编程工具 Bring your own coding agent
Claude Code
OpenAI Codex
Cursor
Kiro
Gemini CLI
+ More
只要 Agent 能读取项目说明，就可以接入这套流程。 Works with any agent that can read project instructions.
↓
Agent Harness 工程流程层Process layer
Scope
Rules
Build
Verify
↓
你的项目仓库Your Repository
问题The Problem

AI 不是不会写代码，而是不知道你的工程流程。 AI agents don't fail because they can't code.

很多 AI 编程失败不是能力问题，而是流程缺失：需求没澄清、边界没确认、架构规则没读取、验证结果说不清。 They fail because they don't know your engineering process: unclear scope, hidden assumptions, skipped project rules and vague verification.

没有 HarnessWithout Harness

AI Agent
直接修改代码Directly changes code
依赖隐藏假设Hidden assumptions
结果不可预测Unpredictable result

使用 Agent HarnessWith Agent Harness

AI Agent
进入工程流程层Goes through process layer
Scope → Rules → Build → VerifyScope → Rules → Build → Verify
结果可审查Reviewable result
架构Architecture

通用流程可复用，项目知识单独沉淀。 Reusable process. Project-specific knowledge.

Agent Harness 将通用流程和项目事实分开：core 负责流程，project 负责当前仓库的规则、约束和上下文。 Agent Harness separates the reusable process layer from local repository knowledge: core defines the workflow, project stores your rules and constraints.

harness/core

通用流程层，可作为 submodule 升级。Generic process layer, upgradeable as a submodule.

Workflow Templates Verification Operations
+

harness/project

当前项目的工程知识，保留在你的仓库里。Local engineering knowledge that stays inside your repository.

Rules Architecture Constraints Risk Paths
能力Capabilities

少而清晰的三层能力。 Three clear layers of capability.

01

工程流程Engineering Flow

让 Agent 在修改代码前明确范围、方案、风险和验证方式。Make agents clarify scope, solution, risk and verification before editing code.

ScopeSolutionBuildVerify
02

项目知识Project Knowledge

沉淀模块边界、架构约束、高风险路径和团队规则。Capture module boundaries, architecture constraints, risk paths and team conventions.

RulesConstraintsArchitecture
03

可复用流程Reusable Process

将 Bugfix、需求、重构、长任务等流程模板化，跨项目复用。Turn bugfixes, features, refactors and long operations into reusable process templates.

TemplatesWorkflowsOperations
真实项目Real Project

已用于真实工具链项目。 Used in a real toolchain project.

Forja

Forja 是一个 C++ 项目构建工具链，包含 VSCode 扩展、独立 CLI 和 AI Skill，支持 Qt/qmake 项目、SDK 项目、Windows MSVC 与 Linux 平台。 Forja is a C++ project build toolchain with a VSCode extension, standalone CLI and AI Skill. It supports Qt/qmake projects, SDK projects, Windows MSVC and Linux.

查看 ForjaView Forja →
VSCode Extension图形化构建、调试、同步GUI build, debug and sync
CLI本地构建和文件同步Local build and file sync
AI Skill为 AI 助手提供构建能力Build capability for AI assistants
Cross-platformWindows MSVC · Linux GCC
Prompt 是临时的。
流程可以规模化。 Prompts are temporary.
Processes scale.
一个 Prompt 解决一个任务，一套流程解决上千个任务。 A prompt solves one task. A process solves thousands.
快速开始Quick Start

三步接入你的仓库。 Add it to your repository in three steps.

1
添加 HarnessAdd Harness以 submodule 方式添加通用流程层。Add the reusable process layer as a submodule.
2
生成项目画像Generate Project Profile让 AI 读取 ONBOARD.md 并生成项目规则。Ask your agent to read ONBOARD.md and generate project rules.
3
按流程开发Start Building让 Agent 按 Scope、Rules、Build、Verify 交付。Make the agent deliver through Scope, Rules, Build and Verify.
git submodule add https://github.com/jxiaow/agent-harness.git harness/core
mkdir -p harness/project/rules

# Then ask your AI coding agent:
Read harness/core/ONBOARD.md
Generate the project profile and rules.