Anthropic's own AI team says one well-equipped agent often beats a crew of subagents, and it published a guide on when to fold them back in.

Anthropic's Applied AI team published a guide on Wednesday, Oct 22, called "When to fold subagents into your main agent loop," and it became one of the most shared AI reads on X this Saturday (Claude, widely shared post).

A subagent is a helper AI that the main agent hands a piece of work to, often with its own instructions and tools. A skill is a packaged set of instructions and tools that one agent loads only when a task calls for it. Anthropic says that across several enterprise deployments, a single agent with skills frequently beat both a single do-everything prompt and a subagent setup on quality, often at lower cost and wait time per task, and one customer saw roughly 2x faster runs.

Subagents still make sense when a long task is clearly getting worse from too much context, when separate jobs can run in parallel with a fresh start, or when one agent struggles to pick between tools or follow conflicting instructions. The guide also warns against designs that copy a company's org chart, and suggests giving the agent short descriptions of each skill and tool and loading the full details only when needed. It notes that tool choice tends to slip once an agent has more than 30 to 50 tools, and that Anthropic cut Claude Code's system prompt for its most advanced models by over 80%.

Its advice is to fold in one subagent at a time and compare results on a small test set run more than once, and it says even 10 good examples are enough to start.

Try it: if you run an agent with several subagents, pick the one that does the least and fold it into the main agent as a skill. Run the same 10 tasks a few times before and after, and keep whichever version does better.