Member-only story
A coding agent can look expensive for at least four different reasons. It can talk too much. It can keep shipping the same logs and file dumps back into context. It can carry tool definitions it will never call. Or it can write a small mountain of code for a small change.
Those are different bills. Treating them as one token problem produces the familiar bad result: the agent gets terse, the work gets worse, and the invoice barely moves.
I would start with one question: what enters the next model request that did not need to? Find that answer before swapping models, adding a proxy, or forcing a short-response prompt.
I am Caspar Bannink, a solo founder building HomeScout, an unrelated rental-search product. This is a builder note about coding-agent cost control, not a HomeScout product announcement. I publish technical field notes at CasparAI on Medium and keep my personal profile at LinkedIn.
The 60% number needs an asterisk #
The open-source token-efficiency scene is suddenly full of dramatic percentages. A project reports 65% fewer output tokens. Another reports 47% fewer request tokens. Another says 26% fewer tokens on a controlled refactor. A fourth reports 20% lower cost by making the agent write less…