Show HN: The Subagent Tax (systima.ai)

Claude Code has revealed findings about the performance and cost implications of using subagents for AI tasks, showing they can significantly increase input token usage compared to sequential processing. In tests across three model families, subagents used 2.6x to 5.9x more tokens, and no speed advantages were observed, with one configuration costing nearly five times more than a straightforward sequential run. This indicates that while subagents may seem like a solution for parallel processing, they incur substantial overhead due to factors like fresh agent loops, duplicate system prompts, and inefficient cache usage. The implications for the AI/ML community are crucial, as the findings highlight the hidden costs associated with subagent orchestration. This performance overhead could lead to inefficiencies and increased expenses when deploying AI models, particularly when using more expensive model families. Notably, Claude's built-in Explore agents now inherit the parent model by default, which can compound the cost if a less expensive model isn't specified. As AI developers seek to optimize resource utilization, these insights encourage careful consideration of agent architectures and the selection of model families to manage expenses effectively.