80,000-agent benchmark: each idle agent costs ~100KB, mostly conversation history
AIFlow_ML · x · 2026-10-25
A hands-on run of the AWS agent benchmark: 80,000 agents × 20 turns each (3.2M completions) simulated on a single thread of an AMD EPYC 9R14, using in-memory stores and a simulated model.
Key findings:
- 100 KB per idle agent, with most of that being conversation history
- Longer histories mean fewer agents per GB of memory
Sources and an Apache-2.0 licensed reproduction script are linked, so you can verify it yourself. Useful reference for estimating memory costs of large-scale agent systems.
Related event: Benchmark runs 80,000 agents on one thread, idle agent uses ~100KB(2 posts)→
More from coding & agent
- Hutash: a two-person team's all-in-one local AI platform now lost in scope — kathi7 · 2026-10-25
- Team ditches Slack for a custom chat app where humans and AI agents work side by side — dee_hw · 2026-10-25
- Musk: Grok Bot can now call businesses for you and summarize the conversation — elonmusk · 2026-10-25
- One prompt let Claude debloat an Android TV over ADB and rebuild a minimal home screen — PrajwalTomar_ · 2026-10-25
- User points Claude at his Samsung TV, spends 11 hours building an ad-free Tizen launcher — PrajwalTomar_ · 2026-10-25
- Jev-Mem paper: a fast scorer makes agent memory building 6.6x faster on LoCoMo — rohanpaul_ai · 2026-10-25