Multi-Agent Companies

A pattern where a set of AI agents is organised as a company org chart — CEO, engineers, designers, marketers, QA — each with specialised roles, instructions, tools, and budgets. Each agent reports to another agent; the CEO (or equivalent) delegates work and reports back to the human board.

Key points

• The org-chart metaphor is load-bearing. It gives each agent a clear identity (persona, responsibilities, reporting line), which improves the quality and coherence of its outputs compared to a generic “assistant” ^[src-001].

• Paperclip ships company templates (GStack, Superpowers, Agency Agents, scientific research teams). One template comes with 48 pre-built agents, each with named roles and specialised skills. This is the equivalent of “acquiring a company for the team” in traditional business ^[src-001].

• Each agent can run on a different model and provider via adapter config. Typical pattern: heavy reasoning roles (CEO, architect) run on Sonnet/Opus; execution roles (designer, marketer) run on Haiku or cheaper models ^[src-001].

• Hiring decisions are themselves delegated to agents. In Paperclip, the CEO agent requests permission to hire a new agent, the board (human) approves, and the new agent is onboarded automatically with instruction files and default skills ^[src-001].

• Approval gates are configurable. By default, new hires require board approval; turning this off lets the CEO continuously grow the team without human intervention ^[src-001].

Related entities

• Paperclip — the primary platform implementing this pattern
• OpenClaw — earlier tool, closer to single-agent model

Related concepts

• Agent Orchestration
• Zero-Human Companies
• Heartbeats (Agent Wake-ups)
• Agent Budget Controls

Source references

• ^[src-001] Nate Herk — “Claude Code + Paperclip Just Destroyed OpenClaw” (2026-03-28)