Empire Architecture

Problem It Solves

A single agent often mixes retrieval, computation, drafting, review, and safety checks into one prompt, making context grow and failure hard to isolate. Empire Architecture uses a Three Excellencies and Nine Ministers metaphor to split work across agent tiers: the chancellor routes and schedules tasks, advisory/executive agents perform specialized work, the Hanlin court handles knowledge retrieval, and the Embroidered Uniform Guard performs audit.

How It Works

• Tag routing: agents carry tags, and the chancellor selects nodes by keywords, reducing unrelated prompt load.
• Model tiers: decision nodes use pro models, while execution and audit nodes can use flash models to control token usage.
• Task queue: supports priority, a 90s timeout, automatic retries with exponential backoff, and circuit breaking after repeated failures.
• Knowledge and memory: supports local RAG, Feishu, Notion, Tencent Cloud knowledge engines, and other sources; agent memory keeps recent context and high-importance items.
• Safety controls: pre-task sensitive keyword checks, post-task audit, violation classification, and voting.

Limits

It fits multi-step, multi-role workflows such as weather analysis, research, report drafting, and code review. It is less useful for one-off simple questions or strictly real-time tasks. If knowledge-source permissions, model settings, or agent tags are incomplete, routing quality and audit results may degrade.