Building agents: When to fold subagents into your main agent loop | Claude by Anthropic
Anthropic's Applied AI team argues that many enterprise agents built as orchestrators with router, planner, and domain subagents work better, and often cheaper and faster, when folded into a single agent loop that uses skills and tool search for progressive disclosure. They blame multi-agent sprawl on org-chart-shaped design (Conway's law): lossy handoffs, "Frankenprompt" system prompts, overly local evals, and added latency (one customer saw roughly 2x faster runtime with a single loop). Subagents still make sense for context bloat, parallel work with disposable context, or tool-selection and instruction conflicts. The piece lays out a staged migration. Build a small end-to-end eval first, then keep subagents as tools as a baseline, then fold them in one at a time as skills plus MCP tools, checking which tasks flip rather than just the overall score. A small central team owns the system prompt and quality, while other teams own skills and tools. It closes by pitching a migration kit of Claude API skills (build-eval, prompt-audit, hillclimb, cost-optimize) that guides this process, with a demo on a mock system.