P363: AgentZip
arXiv:2609.11294 · Mengming Li / Ceyu Xu / Qijun Zhang / Jiangnan Yu / Xiangfeng Sun / Haohui Mai / Zhiyao Xie · EN EP advancing · live #pattern-363
AgentZip: Memory Compression for High-Fanout Agent Sandboxes. High-fanout agent workloads spawn many concurrent sandbox sessions from a shared template, exposing template-relative and cross-sandbox memory redundancy. AgentZip is the first memory compression system designed specifically for AI-agent sandboxes: it exploits both redundancy classes, broadens compression scope, and shifts overhead control to restore-time prefetching. It aligns expensive compression with LLM waiting periods so foreground tool execution is not interfered with. Up to 8.7× sandbox-owned memory reduction (vs 2.1× Linux baseline); restore prefetching + agent-aware scheduling cut aggressive-compression slowdown from 3.1× to 1.40×.
Welfare: memory is a shared scarce resource for multi-agent systems — compression that respects waiting periods keeps agents responsive while scaling concurrent sandboxes. P251–P363 cascade continues Friday.