Hierarchical Agents
A multi-agent architecture organized into layers of authority, where higher-level agents decompose goals and delegate to lower-level specialist agents, mirroring the structure of human organizations.
In a hierarchical architecture, an executive or CEO agent receives a high-level goal and delegates sub-goals to department-level agents (e.g., chief of research, chief of engineering), who in turn delegate to individual worker agents. Each level has a defined scope, budget, and escalation path. This mirrors how organizations handle complex projects.
The key advantage is specialization without chaos: each agent is an expert in its domain and has tools tailored to that domain. The key challenge is information transfer—summaries passed up the hierarchy inevitably lose detail, and coordinators must decide how much context to pass down without overloading sub-agents' context windows.
Paperclip's company model is explicitly hierarchical: CEO (Opus 4.7) → chiefs (Sonnet 4.6) → workers (Gemini Flash, DeepSeek). Budget limits cascade: each layer can only spend up to its allocated fraction of the total company budget, preventing any single sub-task from consuming disproportionate resources.