How to optimize token efficiency in agentic systems

What this post added

This post details how to optimize token efficiency in agentic systems by treating it as a workflow-design problem rather than a prompt-editing problem. It identifies five key areas of token waste: retrieval noise, bloated working context, unstructured memory, over-orchestration, and unbounded loops. The post proposes solutions including measuring token usage by workflow stage, reducing context before it reaches the model through selective retrieval and context engineering, replacing raw transcripts with structured state, routing tasks by complexity, designing loops with clear stopping rules, and reusing proven plans and stable context. It also outlines key metrics for evaluating token efficiency and common mistakes teams make.