For a GraphQL subagent, Shopify reports that Nemotron 3.5 Lightning reached 74.5 on its internal judge after six training loops, compared with 73 for the frontier-model baseline. In a production-traffic replay, median response time was 0.87 seconds versus 6.50 seconds for the frontier model’s API, with serving cost about 90% lower. The company also reports that its self-healing process repaired 59,535 GraphQL conversations over 86 nightly runs, or 69% of the failures it attempted. These are Shopify-reported results from its own workload and evaluation setup.
How Shopify turns open models into task experts
In one minute
Shopify says its Sidekick commerce assistant delegates focused work to subagents, including open models that the company post-trains and serves itself.
What changed: Shopify details Sidekick’s use of self-post-trained open models as task-specific subagents, and how NVIDIA tools and GPUs support their development and serving.