Multi-Agent Patterns

Multi-Agent Patterns is an advanced skill for designing distributed language model systems that overcome single-agent limitations by partitioning work across multiple agents with isolated context windows. When tasks grow complex, single agents face inherent ceilings due to context window constraints and degradation from lost-in-middle effects. Multi-agent systems solve this by distributing work to maintain fresh contexts for each subtask.

The skill covers three primary architectural patterns. Supervisor/Orchestrator uses a central agent that delegates to specialists and aggregates results, providing strict control but creating potential bottlenecks. Peer-to-Peer/Swarm enables agents to communicate directly via explicit handoff protocols, eliminating single points of failure but increasing coordination complexity. Hierarchical organizes agents into strategy, planning, and execution layers, mirroring organizational structures but introducing inter-layer coordination overhead.

Token economics reality is addressed honestly: multi-agent systems consume substantially more tokens than single-agent approaches—approximately 15x baseline for complex coordination. However, research indicates that upgrading to better models often provides larger performance gains than doubling token budgets.

Consensus mechanisms go beyond simple voting (which treats hallucinations equally) to include weighted voting based on confidence, debate protocols requiring mutual critique, and trigger-based interventions detecting stalled progress or agent mimicry. Critical failure modes and mitigations are covered: supervisor bottleneck (use output schemas and checkpointing), coordination overhead (minimize communication through clear protocols), divergence (define objective boundaries and convergence checks), and error propagation (validate outputs before downstream passing).

The core design principle: sub-agents exist primarily to partition context, not to simulate organizational roles. Each agent operates in a clean context focused on its specific subtask. Integrates with LangGraph, AutoGen, and CrewAI frameworks.