📄 中文摘要
多智能体 LLM 协调的理念看似简单，通过同时运行 Claude、Codex 和 Gemini，将任务路由到最适合处理的模型。然而，在经过 86 次会话后，发现同样的安全漏洞出现了三次，TypeScript 配置在每次会话中都被忽视，API 额度在一天内耗尽。Claude Code 多智能体工作流中，必须显式注入上下文，代理之间没有隐式共享。发现的错误必须立即提交到代码中，而不是留待后用。提示约束越紧，输出越稳定。
构建多智能体 LLM 协调器与 Claude Code：86 次艰难的经验教训
📄 English Summary
Building a Multi-Agent LLM Orchestrator with Claude Code: 86 Sessions of Hard-Won Lessons
The concept of multi-agent LLM orchestration appears straightforward: run Claude, Codex, and Gemini simultaneously and route tasks to the model best suited for each. However, after 86 sessions, recurring issues emerged, including the same security bug surfacing three times, consistent neglect of TypeScript configuration, and API credits depleting within a single day. In Claude Code multi-agent workflows, context must be explicitly injected, as there is no implicit sharing between agents. Discovered bugs should be committed to code immediately rather than deferred. Tighter prompt constraints lead to more stable outputs.