为长篇写作项目构建一个由人工智能协调的出版工作流程
2 分•作者: tmuhlestein•2 个月前
我一直在尝试一种“编排优先”的方法，而不是依赖单个模型来处理整个工作流程。
这个项目最初是为了探索一个真实的客户场景，并在此过程中发展成为一个涉及多个专业代理、MCP/A2A 服务、评估循环和支持基础设施的长篇出版工作流。
最终，该系统处理了 260 亿个 token，并发展到包含 25 个代理和工具、30 个代理技能、27 个打包技能、22 个项目、12 个 MCP/A2A 原生服务、8 个全栈服务（API + UI + MCP + A2A）、318 个 PR 和 423 个 commit。
作为 CTO，我的目标是在要求我的工程组织以这种新方式进行构建之前，自己深入研究一个真实的客户场景。
项目地址：https://theisaiahchronicles.com
很想听听其他人是如何处理编排和多代理系统的。
查看原文
I've been experimenting with an orchestration-first approach instead of relying on a single model to handle an entire workflow.<p>The project started as a way to explore a real customer scenario, and along the way it evolved into a long-form publishing workflow involving multiple specialized agents, MCP/A2A services, evaluation loops, and supporting infrastructure.<p>By the end, the system had processed 26 billion tokens and grown to include 25 agents and tools, 30 agent skills, 27 packaged skills, 22 projects, 12 MCP/A2A-native services, 8 full-stack services (API + UI + MCP + A2A), 318 PRs, and 423 commits.<p>My goal as CTO was to go deep into a real customer scenario myself before asking my engineering organization to build in this new way.<p>The project is here: https://theisaiahchronicles.com<p>Would love to hear how others are approaching orchestration and multi-agent systems.