Yang Haoran, a senior technical expert at Alibaba Cloud Intelligence Group, mentioned in a roundtable discussion that workloads in enterprises used to fall roughly into two categories: long-running microservices and offline Jobs with clear lifecycles. After years of development, these two workload types already have relatively mature scheduling and governance systems.
But the way Agents run is clearly more random. What tool it calls next and how many rounds it runs are largely determined dynamically by the model; a large number of Agents still execute intermittently, waiting for model inference during a task and also waiting for Subagents to return results.
What happens to resources while waiting? If CPU and memory remain occupied the whole time, users are paying for waiting time; if the environment is destroyed directly, the task state is gone. The original scheduling approach, which was better suited to long-running instances, becomes less appropriate here.
So Alibaba Cloud this time built both rapid creation and sleep recovery into Agent Sandbox. According to its published figures, Agent Sandbox can create 100,000 sandboxes per minute, with template-based creation and cold-start P99 under 180 milliseconds; after resource pool preheating, warm-start P99 is under 20 milliseconds; and deep-sleep wake-up P99 is under 600 milliseconds. These figures are all unilaterally published by Alibaba Cloud, and the source did not provide third-party verification.