FULL STORY
AWS Open-Sources Infinity Runtime, 80K-Agent Stress Test Follows
AWS and UC Berkeley open-sourced the Infinity agent runtime at AGNTCon. A follow-up stress test ran 80,000 agents with 3.2M model calls on a single EPYC core, showing idle agents use only ~100KB of memory.
2026-10-24 ~ 2026-10-25 · 2 episodes · 5 posts
Episode 1 · AWS Open-Sources Infinity Runtime for Massive Agent Concurrency (2026-10-24, 3 posts)
AWS and UC Berkeley open-sourced Infinity, a Rust runtime that idles waiting agents to claim 75,000 concurrent agents in 8GB, though benchmarks rely on simulated models and the Raspberry Pi figure is an extrapolation.
- Infinity: a Rust agent runtime that fits 75k agents in a Raspberry Pi's RAM — ShadajL · 2026-10-24
- AWS open-sources Infinity: a runtime that sleeps idle agents for extreme concurrency — AIFlow_ML · 2026-10-25
- AWS Infinity under the hood: the 75,000-agents-on-a-Pi number is math, not a Pi test — AIFlow_ML · 2026-10-25
Episode 2 · Benchmark runs 80,000 agents on one thread, idle agent uses ~100KB (2026-10-25, 2 posts)
An AWS-based benchmark simulated 80,000 agents with 3.2 million model calls on a single AMD EPYC thread, showing each idle agent consumes only about 100KB. The key design—tools returning immediately with results pushed asynchronously to agent queues—is seen as directly reusable in any agent stack.
- 80,000-agent benchmark: each idle agent costs ~100KB, mostly conversation history — AIFlow_ML · 2026-10-25
- 80,000-agent stress test: async tools push results to queue, ~100KB per idle agent on one EPYC thread — AIFlow_ML · 2026-10-25