# Shopify says continual training cut its GraphQL agent's serving cost 96%

_Published Tuesday, September 22, 2026 at 6:19 PM EDT · AI, Products · Latest · Tier 2 — Notable_

Shopify says a daily continual-learning system enabled its specialized GraphQL agent to surpass its frontier-model baseline while reducing estimated serving costs by 96%. The production agent handles as many as 2,000 requests per minute, translating merchant questions into queries against Shopify's administrative interface.

The system turns low-scoring, anonymized conversations into repaired training examples, then applies supervised fine-tuning and reinforcement learning. It also compresses a roughly 6,000-token system prompt to about 1,500 learned tokens. The performance and cost comparisons are Shopify's estimates based on its own evaluation system.

## Sources

- [PyTorch](https://pytorch.org/blog/how-shopify-built-a-continual-learning-loop-with-pytorch-and-vllm/)

---
Canonical: https://techandbusiness.org/newswire/ZdyiwADG-QXXx3_fzusRJd
Published: 2026-09-22T22:19:50.560Z
Story chronology: 2026-09-22T13:45:18.000Z
Retrieved: 2026-09-23T00:37:55.846Z
Publisher: Tech & Business (techandbusiness.org)