Jev parallel tool calls and context trimming: 10x faster, 90% cheaper
Adding Jev parallel calls with output caps and OpenAI dots context trimming to a Sonnet 5.5 agent cut median time on 10 tasks from 60s to 6s and cost from $0.30 to $0.03, though only when the calls are independent.
ONE AGENT, 10 ARMS, ONE MERGE FILE - SONNET 5.5 + JEV + OPENAI DOTS RUNS 10X FASTER AND 90% CHEAPER one agent on claude-sonnet-5.5, three changes bolted onto it. nothing retrained, nothing rewritten → jev stops sending tool calls one after another - ten calls fire together, 60s folds into 6s, same task → dots trims what reaches the agent - 40,000 tokens attached to the task drops to 6,000 before any of it gets read → jev caps where the answer stops - 4,000 tokens out becomes 800 → the bill falls in three puffs - $0.300 → $0.180 on the model swap alone, → $0.102 on the context cut, → $0.048 on the capped answer, $0.030 left standing → median across the last 10 tasks: 6.0s, down from 60s, same tokens used → 1,000 tasks a day: $9,000 a month drops to $900 the part nobody puts on the headline card: this only works when the ten calls don't need each other. chain them - each one waiting on the last - and the fold undoes itself, 60s stays 60s, the 10x is gone it's not free on accuracy either. evals went from 226 to 218 out of 240 after the trim. the speed cost something, it just didn't cost money most people show you the $0.030 and stop there. the dominoes chart under it is the part that tells you when $0.030 doesn't apply to your workload $0 extra infra - jev and dots bolt onto an agent you already have, the only thing that changed is how the calls leave and what gets read full breakdown, all 4 cards, is in the video which of your workloads actually fans out like that, and which one is secretly 10 chained dominoes wearing a parallel label?
This post quotes another post: x.com/sprytixl/status/2107763347097534680