Skip to main content

​
The question

Why not let the main agent investigate everything?

​
The answer

Because exploration pollutes the context used to integrate the fix.

​
Without subagents

Auth exploration: 20 files Test logs: 8,000 lines Dependency dead ends User conversation Final reasoning

​
With subagents

Auth → finding Tests → root cause Deps → constraints Main context receives concise results.
  • +90.2% — task success vs. single agent — Anthropic’s own research system
  • ~15× — tokens a full multi-agent run uses vs. one chat turn
  • 80% — of that performance gap explained by token spend alone
These numbers are from Anthropic’s own multi-agent research system (Opus lead + Sonnet subagents on breadth-first search tasks) — not Claude Code specifically. Anthropic explicitly notes the pattern is worse for tightly-coupled work like most coding, which is exactly why the next question gates delegation instead of applying it everywhere.

​
Source

anthropic.com/engineering — built-multi-agent-research-system
Design decision: Delegate isolated exploration, not simply “large tasks.”

​
Continue the walkthrough

The delegation gate

Continue to topic 10 of 14.