Claude Subagents vs. Agent Teams

Claude offers two distinct multi-agent paradigms: sub-agents and agent teams. Sub-agents are isolated, fire-and-forget workers that run in their own context windows, report only final results to a parent, and cannot communicate with each other. Agent teams are long-running, collaborative instances with shared state, peer-to-peer communication, and dependency tracking via a shared task list. The key design principle is context-centric decomposition rather than role-based splitting. Five orchestration patterns are covered: prompt chaining, routing, parallelization, orchestrator-worker, and evaluator-optimizer. The post also warns against over-engineering with multi-agent systems and identifies three common failure modes: vague task descriptions, false verification, and runaway token costs.