Multi-Agent Hierarchy Decision Framework

A decision framework for the most consequential architecture question in multi-agent system design: should this capability be a skill (tool call) or a sub-agent?

Based on synthesis of 50+ sources from 2025-2026 across CrewAI, AutoGen, LangGraph, Magentic-One, Claude Agent SDK, Google A2A, MetaGPT, and production case studies. Core finding: deep agent hierarchies (3+ tiers) fail in production 41-87% of the time. A flat orchestrator with skilled specialists handles 90% of real-world requirements.

The decision matrix uses 7 signals: task determinism, step count, error recovery complexity, latency requirements, token cost tolerance, iteration need, and context sharing requirements. Also covers the 'agent as a tool' pattern, which wraps a complex agent call as a skill and provides sub-agent capability without architectural commitment.

What is included: the full framework document, the 7-signal decision matrix, failure mode analysis from production case studies, the agent-as-a-tool wrapper pattern with code example, and the T2+tools formula.

Best for anyone designing a multi-agent system before they build it, or restructuring one that has become brittle.