How devs are stopping multi-step agents from blowing up token costs
useralltime_835 · reddit · 2026-11-07
A Reddit poster shares practical tactics for taming token costs in multi-step agents: a 3-step agent ballooned to 15 tool calls per task because each step re-sent the full history. Trimming context to only what the next step needs, summarizing older turns, and caching the system prompt cut the bill by more than half with no quality drop. The thread invites other approaches to keeping agent runs lean.
More from coding & agent
- Stardock's draginol: AI agent workflows blur lines between human collaborators and agents — draginol · 2026-11-07
- draginol shows his modern AI dev workflow: code comes much later — draginol · 2026-11-07
- How an AI agent snagged an iPhone Duo as restocks vanished in 14 seconds — jesselyu · 2026-11-07
- Many companies are rewriting Git for agentic workflows — most won't survive GitHub's own fix — arpit_bhayani · 2026-11-07
- Researcher: We used to code better than AI — now the best move is to get out of its way — Yuchenj_UW · 2026-11-07
- GraphRAGProv: provenance polynomials expose wrong-evidence errors in correct GraphRAG answers — JeremyCMorgan · 2026-11-07