Anthropic research finds Claude Code subagents underperform sequential tasks
The author examines when to use multiple coding agent subagents versus a single session in Claude Code, arguing that subagents fail on sequential tasks like refactoring but succeed on parallel searches, distinct modules, or independent code reviews. Drawing on research from Anthropic and Cognition, the author explains that agent teams often appear more effective simply because they consume significantly more tokens. The post recommends running a control comparison with equal budgets to determine whether a subagent setup genuinely outperforms a single session. Open article →