TIL : A single command saves 90% of my tokens
inference-costworkflowtooling
rtk find — a proxy that filters find output before it hits the agent’s
context. Across 908 tracked commands, this single command saved
5.6M tokens out of 6.2M total (90.3%).
The problem: find in a project returns massive file trees
(node_modules, vendor, .git). The agent ingests them entirely. RTK intercepts
the command and filters irrelevant paths before they reach the context.
RTK Token Savings
rtk find 76 invocations 5.6M tokens saved 70.4% efficiency
rtk ls 159 invocations 27K tokens saved 59.6% efficiency
rtk read 115 invocations 15K tokens saved 13.0% efficiency
The insight: in an agentic workflow, the most cost-effective optimization isn’t switching models — it’s preventing noise from entering the context. A single well-placed filter does more than any multi-model routing setup.
Further reading