Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-08 · high
The trick loads each tool's schema only when the model reaches for it, so adding more tools stops taxing every turn.
Why now: Dated to July 8, the 95% figure quantifies a schema-loading cost that's mostly gone unmeasured until now.
MCP agents pay a token tax for every tool schema loaded each turn, whether it's used or not, per an arXiv paper on tool-calling costs. Untouched tools still cost you. That tax scales with tool count, setting a real ceiling on how many tools you can add before every turn gets expensive. The paper's fix, deferring each schema until the model actually reaches for that tool, cut tool tokens per turn by 95% against loading everything upfront.
The paper doesn't say how gating handles a tool the model hasn't called yet but needs mid-conversation. If a schema only loads once the model reaches for it, there's a chicken-and-egg problem for tools it doesn't know exist.
The tool list was never the real cost. The schema was. I'd bet teams pruning their MCP servers down to a handful of tools to save tokens are fixing the wrong variable. Gate the schema instead, and an agent with dozens of tools should cost close to what one with a couple costs.
Each link below shares sources, entities, or timing with this story.
Shared entity: Dynamic / Same source domain / Shared topic / Earlier coverage / Tension
Both cover Dynamic; reported by the same outlet (arxiv.org); overlapping topics (cost, dynamic).
Shared entity: Dynamic / Same source domain / Earlier coverage
Both cover Dynamic; reported by the same outlet (arxiv.org); earlier Dynamic coverage from 2026-07-07.
Shared entity: Dynamic / Shared topic / Earlier coverage
Both cover Dynamic; overlapping topics (dynamic, each); earlier Dynamic coverage from 2026-06-09.