Anthropic says fold subagents back into one agent loop unless evals show they earn their place
In an Oct 22 guide, Anthropic's Applied AI team says a single agent loop with on-demand skills and MCP tools often beat router-plus-specialist designs on quality, cost and latency across enterprise deployments, so keep a subagent only for measured context bloat, parallel throwaway work, tool-choice trouble or direct user contact. The Apache-2.0 single-agent-loop-kit walks you through it: build a ~10-example eval, fold one subagent at a time, and keep any whose replacement scores worse.
