Token Drift: Why AI Agents Slow Down and Cost More Over Time

Token drift describes how an AI agent's effective input context grows as it accumulates conversation history, tool outputs, and retrieved documents, causing longer response times and higher costs per turn. In a 20-turn session, input tokens can balloon from 1,200 to 18,400, with cumulative session costs rising quadratically. Developers must treat context as a budgeted resource, not an unlimited transcript, to manage performance and expense.