Does "rtk" skill cut agent tokens by 60–90%? We tested it

JetBrains published a blog post titled Does "rtk" skill cut agent tokens by 60–90%? We tested it, focused on token use when running coding agents with Claude Code. The piece frames rtk as a skill under test rather than an unmeasured claim, and the headline range of 60–90% is the figure they put forward from that evaluation. The write-up was also linked from HN AI Agents, where the thread sat at 3 points and 0 comments at the time of this summary.

The technical angle is token cost for agent loops, not model quality or new model releases. Agent sessions burn tokens on tool output, file context, and repeated back-and-forth; a skill that trims that surface can change how far a fixed budget goes per task. JetBrains positions rtk as something you plug into that Claude Code path and measure against a baseline, with savings expressed as a large cut in tokens consumed rather than as a single micro-benchmark score.

For engineers and builders, token burn is often the binding constraint on long refactors, multi-file edits, and exploratory agent runs. A verified 60–90% reduction, if it holds on real workloads, would stretch the same API spend and context window across more steps before the session hits limits or cost caps. That matters most for teams already running agents in CI-style loops or heavy local coding sessions where every round trip multiplies context.

The market context is the growing stack of agent tooling around Claude Code and similar coding agents, where skills and wrappers compete on cost and reliability as much as on raw capability. JetBrains entering with a measured token-savings story is a product signal as well as a research note: infrastructure around agents is becoming a place vendors try to differentiate. Early HN traction was thin (3 points, no comments), so the claim is public but not yet stress-tested in a broad community thread.

What to watch next is whether independent runs reproduce the 60–90% band outside JetBrains’ own test setup, and on which task shapes (small edits versus large repo exploration) the savings hold or collapse. Also watch how rtk is installed and scoped as a skill in Claude Code workflows, and whether teams adopt it as a default cost control or only for the noisiest agent jobs. The primary source remains the JetBrains post at blog.jetbrains.com/ai/2026/07/rtk-claude-code-token-savings/.