The Orchestrator's Tax (martinfowler.com)

In a recent exploration of multi-agent systems, the author identifies critical insights on how subagents can negatively impact the performance of their orchestrators due to context overload. Instead of merely accelerating task completion, the true value of subagents lies in their ability to protect the orchestrator's working memory by offloading unnecessary reasoning. The study, prompted by a challenging coding session, reveals that while multiple agents can run concurrently, the overhead associated with context management and duplicated understanding can outweigh the benefits of parallelism, particularly when agents are not synchronized in their tasks. The findings suggest that cognitive locality—keeping tasks that require the same mental model together—should guide how agents are deployed. Rather than focusing solely on saving time, effective use of subagents requires setting firm rules for information management and task delegation. This perspective leads to practical recommendations for improving orchestrator efficiency, including minimizing status checks on running agents and avoiding unnecessary agent spawns for overlapping tasks. By refining how subagents operate, the orchestration of multi-agent systems can achieve improved clarity and functionality, ultimately enhancing the overall productivity of AI-driven workflows.