AI Agent Tax: 60x Cost to Re-Derive Routines

AI agents are paying a significant hidden tax when they re-derive routines they've already performed. An open-source tool developer measured this cost on their own machine. Every morning, their computer-use agent solves the same puzzle from scratch, like opening a dashboard and clicking an export button. It repeatedly pays full price for conclusions it reached the day before. A single desktop frame, measuring 1280 by 800 pixels, costs about 1,365 tokens using Anthropic's image token accounting. When you add context and reasoning, one routine step can reach around 2,509 tokens. The Routine Overhead Ratio, or R, measures the difference between tokens a memoryless agent spends re-deriving a routine and tokens to replay that same routine from a compiled plan. The developer found a median R of 60 times. This means the agent pays 60 times more to re-derive a routine than to simply replay it from a stored plan. This "hidden tax" impacts the efficiency and cost-effectiveness of current AI agent systems.