Intel Tests Xeon Sandbox Provisioning for Concurrent AI Agents
A vendor-run CubeSandbox study measured how long three server platforms kept average environment-creation latency below 200 milliseconds as concurrent requests increased.
Edited by Tyronne Panaino
Intel published a sandbox-provisioning study on October 8 that compares two Xeon server systems with one AMD EPYC system under rising numbers of concurrent CubeSandbox create-and-delete requests. Intel reports later crossings of a 200-millisecond average creation-latency threshold for both Xeon configurations, while its disclosed hardware differences and first-party methodology make this a narrow infrastructure result rather than a general processor verdict.
The intended audience is teams operating coding agents and other systems that create short-lived execution environments. Sandbox creation can sit between an agent request and the work it is allowed to perform, so provisioning delay can affect both responsiveness and the number of simultaneous workflows an infrastructure stack can sustain.
What the benchmark measured
Intel used the CubeSandbox cube-bench create/delete workload against a common template. Each sandbox received the equivalent of two CPUs and about two gibibytes of memory. The workload continuously sent creation requests through the CubeSandbox API while aggregate concurrency rose from 100 to 350 requests.
The reported metric was average sandbox-creation latency. Intel treated 200 milliseconds as the service-level threshold for a usable environment, then identified where each system's average crossed that line. This design measures provisioning behavior for one template and request pattern; it does not measure an agent's reasoning quality, task completion rate or the security strength of the sandbox.
Intel reports later threshold crossings on Xeon
According to the Intel sandbox study, the Xeon 6990E+ system stayed below the 200-millisecond average threshold until about 313 concurrent requests. The Xeon 6980P system reached about 277 requests before crossing it, while the AMD EPYC 9965 system crossed at about 246.
Intel translates those crossing points into roughly 27% higher within-threshold concurrency for the 6990E+ configuration and about 12% higher for the 6980P configuration than the tested AMD system. Those are vendor-reported results from testing conducted in September, not independent measurements reproduced across several operators.
The configurations are not interchangeable
The disclosed systems differ in more than processor brand. Intel lists 288 cores for the 6990E+ platform, 128 cores for the 6980P platform and 192 cores for the AMD EPYC 9965 platform. The systems also use different server designs, BIOS and microcode versions, memory speeds, network devices and storage.
That disclosure makes the experiment inspectable, but it also limits simple per-processor conclusions. The result describes the complete configurations Intel tested under one CubeSandbox workload. It does not isolate the contribution of core count, memory, firmware, networking or storage, and it does not provide a cost-normalized comparison.
The article reports average creation latency and threshold crossings. It does not publish an independent reproduction, a tail-latency table or results across multiple sandbox implementations and agent workloads. Infrastructure teams should therefore reproduce the request shape, sandbox size and control-plane behavior that resemble their own deployment before treating the percentages as capacity-planning inputs.
Why provisioning belongs in agent evaluation
Intel's October newsroom index identifies the study as an October 8 data-center publication. The study's broader premise is that modern agents may run code, process files, query databases and call APIs, which makes isolated environments relevant to both security and operations. Virtual machines, containers, microVMs and CubeSandbox are listed as current isolation approaches.
Provisioning speed is only one part of that decision. Intel's own notice says performance varies by configuration and that results may not reflect all later updates. A production evaluation also needs security boundaries, teardown behavior, failure isolation, tail latency, utilization and cost, none of which is resolved by the reported crossing points alone.
Status
Learning. Intel provides the benchmark setup, crossing points and configuration details for the study it ran. Internal confidence is medium because the evidence is first-party, the tested systems differ materially, and no independent reproduction or wider workload set was fetched.
Sources
Update note: Last reviewed 2026-10-12. We will revise this post if independent reproductions or materially different sandbox workloads become available.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.