How to Optimize Token Efficiency in Agentic Systems
Glean, Thursday, July 23rd, 2026
Treating token efficiency in agents as a workflow-design problem, not prompt editing.
This article frames token efficiency in agentic systems as a workflow design challenge rather than a prompt-editing task.
It identifies five major sources of token waste: retrieval noise, bloated working context, unstructured memory, over-orchestration, and unbounded loops.
Recommendations include measuring token usage by workflow stage, reducing context before it reaches the model, replacing raw transcripts with structured state, routing tasks by complexity, and bounding loops.
The key insight is that "better outcome per token" matters more than simply minimizing token count.