What is the 'telephone game' problem in AI orchestration?

This occurs when a supervisor agent incorrectly paraphrases or loses detail from a sub-agent's response. This skill provides direct pass-through patterns (forward_message tools) to maintain data fidelity.

When should I use a multi-agent system instead of a single agent?

Use multi-agent systems when a task's complexity exceeds a single context window, requires parallel processing of independent subtasks, or needs specialized toolsets and system prompts for different domains.

How do multi-agent patterns help with context limits?

They use context isolation to partition work across multiple instances. Each agent operates in a clean window focused on a specific subtask, preventing the 'lost-in-the-middle' effect common in long contexts.

Which frameworks are compatible with these patterns?

These patterns are foundational and apply to major frameworks including LangGraph (graph-based), AutoGen (conversational), and CrewAI (role-based), as well as custom implementations.

Multi-Agent Architecture Patterns

Name: Multi-Agent Architecture Patterns
Author: EricGrill

byEricGrill

•

Data Science & ML

Guides the design and implementation of distributed agent systems to maximize context efficiency and task parallelization.

This skill provides comprehensive guidance for architecting complex multi-agent systems using Claude. It focuses on solving context window limitations through strategic isolation patterns like Supervisor/Orchestrator, Peer-to-Peer Swarms, and Hierarchical layers. By emphasizing context partitioning over simple role-playing, it helps developers build scalable AI workflows that minimize coordination overhead, avoid information loss during agent handoffs, and leverage parallel reasoning for sophisticated enterprise-grade applications.

Key Features

01Mitigates failure modes like supervisor bottlenecks and coordination drift

02Addresses context window management and strategic isolation

03Provides protocols for consensus, weighted voting, and adversarial critique

04Implements Supervisor, Swarm, and Hierarchical architecture patterns

05Optimizes token economics and performance-to-cost ratios

062 GitHub stars

Use Cases

01Building enterprise workflows with strategic, planning, and execution layers

02Scaling agentic capabilities beyond the limits of a single context window

03Designing complex research pipelines requiring parallel document analysis

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add ericgrill/agents-skills-plugins multi-agent-patterns

For use in Claude.ai and ChatGPT

Download Skill