# orch-agents: A Self-developed Agent Orchestration System > A development workflow orchestration system inspired by OpenAI Symphony, integrating GitHub Webhooks and Linear to automatically schedule multi-agent teams for tasks like code review, bug fixing, feature development, etc. - 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm) - 发布时间: 2026-04-18T18:15:01.000Z - 最近活动: 2026-04-18T18:20:41.164Z - 热度: 150.9 - 关键词: AI智能体, GitHub自动化, 代码审查, 工作流编排, Claude Code, Linear, DevOps, CI/CD - 页面链接: https://www.zingnex.cn/en/forum/thread/orch-agents - Canonical: https://www.zingnex.cn/forum/thread/orch-agents - Markdown 来源: floors_fallback --- ## 【Introduction】orch-agents: Core Introduction to the Self-developed Agent Orchestration System orch-agents is a development workflow orchestration system inspired by OpenAI Symphony. By integrating GitHub Webhooks and Linear, it automatically schedules multi-agent teams to handle repetitive tasks such as code review, bug fixing, feature development, and CI failure resolution. Its goal is to reduce engineers' time spent on these tasks so they can focus on more creative work. ## Project Background: Why Do We Need an Agent Orchestration System? In software development, repetitive tasks like code review, bug fixing, and CI failure debugging take up a lot of engineers' time. Inspired by OpenAI Symphony and Linear Agent Interaction Guidelines, the orch-agents project aims to automate these workflows through agent orchestration, building a scalable system that monitors events from code repositories and project management tools, then schedules the appropriate AI agent teams to complete tasks. ## Core Architecture & Approach: Event-Driven + Agent Team Templates ### Event-Driven Architecture The system uses an event-driven design with the following flow: Webhook arrival → signature verification → input cleaning → event standardization → search for WORKFLOW.md → template matching → create agent workspace → sequential execution → quality review → submit results → clean workspace. It supports GitHub and Linear event sources, which are uniformly mapped to the internal IntakeEvent format. ### Agent Team Templates WORKFLOW.md is the core configuration file, defining agent team templates (e.g., tdd-workflow corresponds to coder + tester) and event routing rules (e.g., opening a GitHub PR triggers github-ops, tagging an Issue with 'bug' triggers tdd-workflow). ### Isolated Execution Environment Each agent runs in an independent Git worktree to ensure environment isolation, security boundaries, and rollback capability. ## Key Features & Quality Assurance: Agent Types & Review Mechanisms ### Agent Types The project includes over 18 built-in agents covering various development stages: architect (system design), coder (code implementation), tester (test writing), reviewer (code review), security-architect (security audit), etc. Each agent's role and specifications are defined via Markdown files in the .claude/agents/ directory. ### Quality Review Mechanisms - **Review Gates**: After each agent completes its work, it must pass three reviews: Claude-driven diff review, npm test execution, and sensitive information security scan. - **Fix Loop**: If a review fails, it automatically enters a fix loop (generate suggestions → agent fixes → re-review), with a maximum of 3 attempts. If it fails after that, the task is aborted. ## Integrations & Use Cases: GitHub/Linear Support & Typical Applications ### GitHub Integration - **Webhook Configuration**: The payload URL is https://your-server/webhooks/github, content type is application/json. You need to configure WEBHOOK_SECRET and listen for events like Push, PR, Issues. - **Authentication Mode**: GitHub App is recommended (operations are shown as orch-agents[bot], preventing feedback loops). Personal access tokens are also supported (suitable for development and testing). ### Linear Integration Connect to Linear via configuration files. When a card moves to an active state (unstarted/started), the corresponding workflow is triggered. ### Use Cases - Automated Code Review: After a PR is submitted, the reviewer agent automatically reviews it and posts comments. - Intelligent Bug Fixing: When an Issue is tagged with 'bug', the coder+tester team automatically analyzes the root cause, writes fixes and tests, and submits a PR. - CI Failure Resolution: When CI fails on the main branch, the quick-fix agent automatically locates the problem and pushes a fix. - Feature Development Pipeline: When an Issue is tagged with 'feature', architect → coder → reviewer sequentially complete the process from design to review. ### Tech Stack Runtime: Node.js 22+, Language: TypeScript, AI Engine: Claude Code, Tests: Over 730 cases, Deployment: Supports Docker and Cloudflare Tunnel. ## Limitations & Considerations: Cost, Accuracy & Configuration Complexity ### Cost Considerations Agent calls consume Claude API tokens, which may lead to significant costs for large projects or high-frequency events. ### Review Accuracy AI-generated code may still have logical errors or improper boundary handling; critical code requires final manual review. ### Configuration Complexity The routing rules and conditional judgments in WORKFLOW.md have a certain learning curve. ## Conclusion & Recommendations: Potential of AI-Driven Development & Value of Trying orch-agents demonstrates the great potential of AI agents in software development workflows. Through reasonable orchestration and quality control, it significantly reduces engineers' time spent on repetitive tasks. For teams exploring AI-driven development processes, this is a project worth researching and trying. Its modular design and clear architecture also provide a good reference for building complex agent systems. It is recommended that interested teams deeply understand and practice this system.