← sources

How are you keeping multi step agent runs from blowing up token cost?

source: reddit (reddit:AI_Agents)published 2026-11-07 04:36ingested 2026-11-08 02:35

retained text

A three step agent quietly turned into fifteen tool calls per task because each step re sent the whole history. Trimming the context to only what the next step needs and summarizing older turns cut the bill by more than half with no quality drop. Caching the system prompt helped too. What is your approach to keeping runs lean?

document fetched 2026-11-08 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1xpxlc8/how_are_you_keeping_multi_step_agent_runs_from