How to reproduce performance metrics and scaling results from DGX Spark autonomous agent blog?

Hi NVIDIA team,

I was reading the following blog post:

I’m very interested in the performance and scaling results presented (e.g., agent workflows, multi-agent orchestration, and DGX Spark throughput/latency improvements), and I would like to better understand how to reproduce these results in practice.

Could you please help clarify the following:

  1. Benchmark methodology
  • What specific benchmarks or metrics were used to measure performance (e.g., tokens/sec, latency, throughput, agent task completion time)?
  • Are there standardized benchmarking scripts or tools available?
  1. Reproducibility / setup
  • Is there a reference implementation, sample code, or GitHub repo for these experiments?
  • What exact software stack was used (e.g., NIM, TensorRT-LLM, specific frameworks, orchestration tools)?
  1. Workload configuration
  • How were the autonomous agents configured (number of agents, task types, orchestration method)?
  • Were these single-node DGX Spark tests or distributed setups?
  1. Step-by-step guidance
  • Are there any official guides or best practices to reproduce similar scaling results on DGX Spark?
  • For example, starting from a simple agent workflow and scaling to multi-agent systems.

From what I understand, DGX Spark enables running large models locally and supports building agent-based workflows with NVIDIA AI software stack , but I would like to better understand the practical steps to validate the results shown in the blog.

Thanks in advance for your help!