Let the Running Tool Report Its Own Progress and KV-Cache Eviction Beats Every Published Duration Predictor

arXiv 2609.18849·medium signalXBlueskyLinkedInCopy link

arXiv 2609.18849 (16 Sep 2026) argues serving systems currently guess how long an agent's tool call will run from its name, its history, a pre-declared duration, or engine occupancy, and proves no estimate fixed before the call starts can even rank calls correctly. A census of four public agent corpora finds a readable progress signal in most tool time once the tool is allowed to emit it, either as fraction-of-work-remaining or an end-is-near flag. At the moments a KV-cache decision is made, reported progress is several times to an order of magnitude more accurate than the best published predictors, holds up under environment change, and cuts p90 time-to-first-token in a production engine through a few small hints.