Why AI agents burn through tokens: resent context costs add up each turn

The author argues that an AI agent's output quality often depends less on the model or prompt than on the economics of its context window. Because agents resend their entire working context on every turn, sloppy context is paid for repeatedly until a run ends. Common culprits include tool schemas that ship with every request even when unused, large tool outputs that linger in history, files re-read unnecessarily, stale logs, and verbose reasoning traces. The writer suggests a rough model in which cost scales with context size multiplied by how many turns it survives, which makes small inefficiencies dominant in long agent loops. Bloated context also dilutes the model's attention, so trimming waste can improve both cost and performance. The article recommends profiling the actual context sent to the model as a first step.