Anthropic: When to fold subagents into your main agent loop

Anthropic's Applied AI team recommends considering the folding of routers, planners, and domain subagents into the main agent loop, finding that single-agent designs with skills frequently outperform multi-agent architectures in quality, and often at a lower cost and latency per task in enterprise comparisons.
In a customer support case, a single-agent loop recorded approximately 2x faster runtime compared to a multi-agent harness, while multi-agent setups caused brand inconsistency and latency issues.
Subagents remain useful for context expansion, parallel tasks requiring fresh windows, or when tool selection fails; otherwise, they add unnecessary complexity and 'Frankenprompts'.
The team released a migration kit with tools like /claude-api hillclimb and /claude-api cost-optimize to help teams iteratively fold subagents into skills and tools while maintaining quality via end-to-end evals.

Design Principles

Use a single model and context window with an open-ended tool loop; move domain procedures to skills and hard limits to tools.
Implement progressive disclosure for context and tools to prevent context bloat; Claude Code's system prompt for advanced models was reduced by over 80% using this method.
Central teams should own the system prompt and end-to-end quality, while domain teams contribute via knowledge bases, skills, and MCP tools.

Build small end-to-end evals (starting with 10 golden examples) owned by a central team to measure task completion.
Fold subagents one by one into skills or tools; keep the subagent only if the folded version scores worse.
Monitor for task flips (success/failure changes) rather than just final scores, and run each task at least twice to check consistency.