What they did

The authors built AgentZip, a memory-compression system for high-fanout agent workloads. It represents sandbox memory using redundancy relative to a common template and similarities among pages from different sandboxes, including pages that are related but not identical.

Rather than conservatively deciding which pages to compress, AgentZip compresses any page for which a smaller representation is available and manages restoration costs through prefetching. It also schedules expensive compression during periods when tools are waiting for LLM responses, reducing interference with foreground execution.

Key findings

• AgentZip reduced sandbox-owned memory by up to 8.7× across the evaluated LLM training and inference workloads.
• The comparison Linux configuration achieved a maximum reduction of 2.1×.
• Aggressive compression without the proposed overhead controls could slow execution by as much as 3.1×.
• Restore prefetching and execution-aware scheduling reduced that slowdown to 1.40× while preserving nearly all of the memory savings.

Why it matters

Memory can constrain how many agent sandboxes run concurrently, especially when one task branches into many related sessions. By exploiting redundancy across those sessions and moving compression work into LLM waiting periods, AgentZip offers a systems-level route to higher sandbox density without proportionally increasing physical memory.

Caveats

The reported 8.7× reduction is a maximum across the evaluated workloads, not a guaranteed result for every sandbox configuration. Even with prefetching and scheduling, aggressive compression retained a 1.40× slowdown; broader results across agent frameworks, trajectory diversity, hardware configurations, and latency-sensitive workloads are needed to establish generality.