High‑fanout agent workloads create a severe memory bottleneck because a single task may spawn many concurrent sandbox sessions. These sandboxes are not independent: they originate from a shared template and run related trajectories, exposing substantial template‑relative and cross‑sandbox redundancy. Conventional memory compression mismatches this setting in three ways: it does not exploit similarity across non‑identical pages; it limits what to compress via conservative page selection, incurring page‑fault overhead; and it triggers compression either only under memory pressure or without regard to agent execution phases.
AgentZip is the first memory‑compression system built specifically for AI‑agent sandboxes. It introduces mechanisms that exploit both template‑relative and cross‑sandbox redundancy, broadens the compression scope to any page with a profitable representation, and shifts overhead control from compression‑time page selection to restore‑time prefetching. Moreover, it aligns heavyweight compression with large language model (LLM) waiting periods, avoiding interference with foreground tool execution.
Across LLM training and inference workloads, AgentZip reduces sandbox‑owned memory by up to 8.7×, compared with 2.1× for the default Linux configuration. Restore prefetching and agent‑execution‑aware scheduling cut the slowdown of aggressive compression from as high as 3.1× to 1.40× while retaining nearly all of its memory‑saving benefit.
Review: AgentZip aligns heavy compression with LLM idle periods, achieving substantial memory savings while keeping overhead low.