Multi-Agent Patterns

A 10-pattern taxonomy organises the design space for multi-agent LLM systems. Patterns are characterised along four dimensions: concurrency (sequential vs. parallel), context isolation (shared vs. separate), human participation (observer vs. peer), and turn-taking (structured vs. free-form).

Pattern Taxonomy

P0 — Role Simulation: Single model plays multiple roles. Baseline approach. No true independence.

P1 — Autonomous Loop: Single agent in a self-directed perceive-reason-act cycle. No inter-agent communication.

P2 — Sequential Pipeline: Agents process in fixed order. Output of one becomes input of next. Clear audit trail.

P3 — Orchestrator-Worker: Central orchestrator delegates tasks to specialist workers. Most common enterprise pattern. Single point of coordination.

P4 — Role-Based Crew: Named roles with defined responsibilities (CrewAI model). Structured collaboration with role-specific prompts.

P5 — Group Chat: Open multi-agent discussion (AutoGen model). Flexible but harder to control.

P6 — Handoff: Agents transfer control to each other based on expertise (Swarm pattern). Dynamic routing.

P7 — Personal Assistant + Heartbeat: Persistent agent with scheduled background monitoring. Enables proactive behaviour via heartbeat mechanism.

P8 — Organisational Hierarchy: Agents in management layers (Paperclip model). Mirrors organisational structure.

P9 — Shared Room: Concurrent agents with enforced context isolation and human operator as peer participant. Most aligned with nuclear control room operations. Least developed in current literature.

Characteristic Failures

Each pattern has distinctive failure modes. P0 cannot achieve epistemic-independence. P3 creates single points of failure at the orchestrator. P5 can produce runaway conversations. P9 requires careful context-divergence management.

Nuclear Relevance

Pattern 9 (Shared Room) maps most naturally to nuclear control room operations: multiple specialist agents operating concurrently with enforced separation, a human operator participating as a peer with final authority, and structured delivery-modes for managing information flow. The full Pattern 9 deployment with 7 agents is described in emergency scenarios like the LOCA response (see emergency-response).

Pattern selection is not purely technical — it determines the achievable level of epistemic-independence, the natural human-authority integration points, and the situation-awareness characteristics of the system.