Making agentic token costs visible in production

This article examines how compounding token costs in agentic AI systems are driven by large tool schemas, accumulating session histories, and redundant retrieval loops. To mitigate these expenses, it recommends implementing strategies such as pruning tool catalogs, using history summarization or sliding windows, and deduplicating retrieved data. Ultimately, the author emphasizes that successful cost management requires robust token visibility to identify and alert on specific drivers of usage bloat.