In an agentless architecture, the submitter decides which machine or orchestrator to use for job submission. In more complex situations, such as jobs that depend on other jobs, or design exploration scenarios, this means that they must work out how many jobs to submit and when to submit them to meet the demand of the workflow.

In an agent-based architecture, the system can decide how many extra agents (workers) to launch to meet the demand. The algorithm by which scaling decisions are made is called a scaling strategy. The system will automatically scale up agents to pick up the work and scale down those agents when the work is done. An agent can run multiple tasks so there is not necessarily a 1-1 mapping between the number jobs and tasks run.

The component in Ansys HPC Platform Services which performs this service is called the autoscaler. The autoscaler works in conjunction with Slurm, LSF, PBS, UGE, AWS ParallelCluster, and Kubernetes.

Agents (workers) that are autoscaled are referred to as ephemeral agents (or ephemeral evaluators in the context of Ansys HPC Platform Services).

For a more detailed look at how the autoscaler works, see Autoscalers.