HKUST's AgentZip cuts agent sandbox memory up to 8.7x by exploiting cross-sandbox redundancy
omarsar0 · x · 2026-09-12
HKUST researchers propose AgentZip to tackle the memory bottleneck of running many parallel agent sandboxes for RL or evals.
Problem: concurrent sandboxes start from the same template with related trajectories; 76–96% of pages show template-relative or cross-sandbox redundancy, making memory the capacity limit for high-fanout agent workloads.
Approach:
- Compress pages against the template and sibling sandboxes, including similar-but-not-identical pages
- Perform expensive compression while the agent waits on the LLM
- Prefetch pages at restore time to limit slowdown
Results: sandbox-owned memory drops up to 8.7x vs 2.1x for a Linux configuration.
More from Infra
- DeepSeek v4.1 Flash runs out of the box on six NVIDIA GPUs via vLLM on day 0, AMD lags — woosuk_k · 2026-09-12
- Open-source Plano routes LLM calls by prompt intent, no agent code changes, cutting bills 2x — Roger_M_Taylor · 2026-09-12
- Netflix Engineer Open-Sources Headroom, Cuts Agent Token Use by Up to 95% — Roger_M_Taylor · 2026-09-12
- DeepSeek V4.1 Flash cuts global KV cache to 890 bytes/token, but HBM demand may rise with agent swarms — teortaxesTex · 2026-09-12
- Polymarket Prices AI Bubble Burst at 12% Odds Through End of 2026 — Polymarket · 2026-09-12
- SGLang hits 873 tok/s on DeepSeek V4.1 Flash within 24 hours of launch — BanghuaZ · 2026-09-12