← sources

What’s your agentic token ratio?

source: reddit (reddit:LocalLLaMA)published 2026-09-20 01:03ingested 2026-09-21 02:38

retained text

Despite crazy optimization of my harness, on a long agentic task, is still giving me crazy ratios of 185:1 and I’m constantly fighting with pre fill delay.

What’s your ratios like?

Edit: I’m running Qwen3.8-27B, no caching

document fetched 2026-09-21 00:37 · canonical https://reddit.com/r/LocalLLaMA/comments/1wl2zmt/whats_your_agentic_token_ratio