80,000-agent stress test: async tools push results to queue, ~100KB per idle agent on one EPYC thread
AIFlow_ML · x · 2026-10-25
An agent-stack pattern worth copying: tools return immediately and push their result to the agent's queue later, so a 100ms lookup and a 3-day CI run become the same code, and nothing sits in memory while waiting.
Backing benchmark: 80,000 agents, 20 turns each, 3.2M completions on one thread of an AMD EPYC 9R14 with in-memory stores and a simulated model. Each idle agent uses 100KB, mostly conversation history; longer histories mean fewer agents per GB.
Related event: Benchmark runs 80,000 agents on one thread, idle agent uses ~100KB(2 posts)→
More from coding & agent
- Law firm's Claude rollout goes sideways: the gap is enablement, not tools — KirkNewcombe · 2026-10-25
- Subagents as browser tabs: the mental model for isolated Claude Code context — daniel_mac8 · 2026-10-25
- EC2-Bench: edge-cloud collaboration hits 73% of cloud-only agent performance at 15% of the cost — zibuyu9 · 2026-10-25
- Dev uses an AI agent to install Jellyfin, sideload the TV app, and configure mounts — DanielLockyer · 2026-10-25
- LocalCode: open-source Windows coding desktop for Ollama hits early preview — ChengyiZhang · 2026-10-25
- Kimi K3.1 Pro vs Claude Opus 5.5: 13 agentic coding tasks for 55% of the bill — Prompt Engineering · 2026-10-25