The article argues that routers, planners, and domain subagents can sometimes be folded into one main agent loop, with skills and tools loaded only when needed.
The proposed loop uses one model and one context window: the model plans, makes batches of tool calls, receives raw results, and decides whether to continue, ask the user, or answer. The harness retains turn and cost budgets, write gates, and optional goal or verifier checks. Anthropic says a customer comparison found the single-loop version ran roughly twice as fast, and says it reduced Claude Code's system prompt by more than 80% for advanced models through progressive disclosure.
The guidance does not call for eliminating subagents. It names context-related quality degradation, parallel independent work with disposable context, tool-selection difficulty, and conflicting instructions as cases where subagents remain appropriate. Anthropic recommends shared end-to-end task-completion evaluations, beginning with as few as 10 golden examples, to compare architectures.
The source describes Anthropic's observations across several enterprise deployments and selected customer examples. It does not provide a universal benchmark, full evaluation data, or evidence that a single loop will outperform a multi-agent design in every workload.