Blog
Engineering notes on token efficiency and AI coding agents
Amazon replaced Claude Sonnet with a small decision model for one task and cut its cost 98%. Coding agents have the same job opening. Here is what Amazon did, what RTK pipelines do with a local decision model, and what it saves: 14% to 27% of total cost in our benchmarks, with the losses shown.
Read articleEvery RTK pre-release now goes through 1,000+ differential fuzz cases, a deterministic oracle gate and an advisory decision-model judge. Here is how it works, what it checks, and what comes next.
Read articleRTK v1 is out. The battle-tested CLI compressor becomes a full cost reduction engine: compress more than bash, route work to cheaper models, build your own pipeline, plug in new methods, and share it, in any agent. Upgrading changes nothing.
Read articleA JetBrains study measured RTK at +7.6% more expensive on SkillsBench at low reasoning effort, ±0% at high effort, with no quality change. We reproduce this benchmark.
Read article