The original title is: "Token Drift Explained: Why Your Agent Gets Slower and More Expensive"

Short summary

Token drift occurs when an AI agent's effective input context grows each turn as conversation history, tool outputs, and retrieved documents accumulate, making later calls slower and more expensive. The article explains that cumulative session cost grows quadratically even when per-turn input grows linearly, and identifies six common sources of context bloat beyond just chat history. It recommends treating context as a budgeted resource, tracking token usage per model call, and leveraging prompt caching where available.

• •Token drift: effective input context grows each turn, increasing latency and cost
• •Cumulative session cost grows quadratically because growing transcripts are resent each turn
• •Context should be budgeted: track per-call usage, trim tool payloads, use prompt caching

Generated with AI, which can make mistakes.