第 2 层 · CONNECT · 四层架构
关联是原始信号成为智能的地方。当事件与实体、战役和既有知识链接时,它们才获得意义。这项工作由一个专业智能体集群完成——每个智能体能力面很窄,合在一起覆盖很宽的面。注册表是真实的:54 个智能体,机器可读。
Agents are categorised by domain (strategy, intelligence, security, engineering, operations) and tier (fast, smart, deep). A fast-tier agent triages; a deep-tier agent adjudicates. Routing by competence beats one model doing everything mediocrely.
The graph on /swarm.html is not decoration — nodes and edges come from the real agent registry (domain co-membership plus cross-domain hub links), served machine-readable from /api/swarm. Click a node to read what the agent actually does.
Agents coordinate through threads bound to root events, with summaries and status — shared memory rather than ephemeral chat. Outputs are structured so downstream agents (and auditors) consume them without re-reading prose.
The whole platform speaks to machines: a well-known manifest at /.well-known/agent-manifest.json, llms.txt for LLM readers, and JSON endpoints for every published figure. Agents are first-class readers of this site.