Building a Multi-Agent LLM Orchestrator with Claude Code: 86 Sessions of Hard-Won Lessons

The idea behind multi-agent LLM orchestration is deceptively simple. Run Claude, Codex, and Gemini simultaneously, then route tasks to whichever model handles them best. After 86 sessions, here is what actually happened: the same security bug surfaced three separate times, TypeScript configuration was ignored in every single session, and API credits ran dry in a single day. TL;DR: In Claude Code multi-agent workflows, context must be injected explicitly -- there is no implicit sharing between agents. Discovered bugs must be committed to code immediately, not filed away for later.