wanman.ai：让 AI agents 团队帮助用户运营一人公司
来源: 郭宇（guoyu.eth），X (Twitter) 推文 发布日期: 2026-04-25
🇬🇧 EN: wanman.ai officially launched and went open source — the core idea is using AI agent teams to help users run a one-person company.
Two work modes: 1. Goal-driven autonomous operation: analyze goals → plan tasks → invite AI employees → auto-meeting alignment → creative brainstorming 2. GitHub repo takeover: auto-analyze code repos, continuous optimization/testing, submit PRs
Core features: message-based communication, self-evolving Skills, sandbox isolation, multi-model architecture.
Design philosophy: let humans step back to an observer role. Core architecture already open source on GitHub (chekusu/wanman).
Name comes from Japanese ワンマン (one-man) romanization.
At its core, wanman runs a JSON-RPC supervisor that coordinates multiple agents (CEO, Developer, Operations, Marketing, Feedback, etc.) as real Claude Code or Codex CLI subprocesses, each isolated in independent worktrees and $HOME environments.
The framework is CLI-first, all operations can be scripted, observed, and reproduced via commands and JSON-RPC supervisor.
🇨🇳 中文原文: wanman.ai 宣布开源并正式上线，核心理念是让 AI Agents 团队帮助用户运营一人公司。
两种工作模式： 1. 从故事目标自治运行：分析目标 → 规划任务 → 邀请 AI 员工 → 自动开会对齐 → 创意发散 2. GitHub 仓库接管：自动分析代码仓库、持续优化测试、提交 PR
核心特点：消息通信、自进化 Skill、沙箱隔离、多模型架构。
设计哲学是让人类退居二线仅作观察者，核心架构已在 GitHub 开源（chekusu/wanman）。
命名来源于日语 ワンマン（一人）的罗马音。
其核心是一个 JSON-RPC supervisor，编排多个智能体（CEO、开发者、运维、市场、反馈等）作为真实的 Claude Code 或 Codex CLI 子进程运行，每个智能体隔离在独立的工作区和 $HOME 环境中。
该框架是 CLI 优先的，所有操作都可以通过命令和 JSON-RPC supervisor 进行脚本化、可观测和可复现。