Agent teams debate dynamic tool routing: 40 schemas per turn is burning tokens

Abject_Housing7279 · reddit · 2026-09-11

A team's agent sends forty tool schemas — verbose JSON descriptions, duplicated rules, examples, parameter notes — on every turn, when their router could rule out thirty-eight before the model sees a token. They're testing dynamic tool routing and schema pruning but face tricky tradeoffs: too narrow a gate hurts recall, too broad inflates input tokens, weakens prefix caching, and drifts time-to-first-token upward. They also distrust cost-per-call since cheap calls needing retries are still expensive tasks. The post asks how others decide which tools enter the prompt each turn and what metrics prove the router helps.

Original post →

More from coding & agent

coding & agent channel →