The Orchestrator's Tax
The article "The Orchestrator's Tax," published on July 28, 2026, by Rahul Garg, a Principal Engineer at Thoughtworks, examines the role of subagents within long-running multi-agent systems. Garg contends that the primary justification for subagents is not merely the time saved or parallel execution, but rather their capacity to safeguard the orchestrator's working memory. He highlights that every token within an orchestrator's context competes for its attention, establishing the orchestrator's working memory as a critical and scarce resource. Consequently, the true value of a subagent is defined by its ability to keep information out of the orchestrator's context, rather than its operational speed. The author advocates for treating subagents as a mechanism to protect the orchestrator's working memory by offloading reasoning that the orchestrator does not need to persistently hold. This approach, which the article terms "Cognitive Locality," shifts the understanding of parallelism's purpose. Implementing this effectively involves providing the orchestrator with explicit ground rules that dictate when and how to delegate tasks, moving from transient sessions to established standing rules. The article, which stems from a real incident, is presented as exploratory work, concluding with more open questions than settled answers regarding these architectural considerations.
Developers can optimize multi-agent systems by designing subagents to protect the orchestrator's working memory through strategic delegation.