Shopify CTO unpacks AI-native engineering at scale

In this Latent Space episode, Shopify CTO Mikhail Parakhin describes a company that has moved well past surface-level AI integration. He points to a December 2025 inflection point when model quality shifted enough to change behavior across the organization, ultimately driving nearly everyone at Shopify to use AI tools on a daily basis. Rather than celebrating raw throughput, Parakhin stresses that token budgets only matter when paired with strong critique and review loops — and that the anti-pattern is spinning up too many parallel agents that don't communicate with each other, which he says burns tokens inefficiently compared to fewer, well-coordinated agents with a dedicated critique step.

Tangle is described as a reproducible workflow system for ML and data experimentation.

The episode goes deep on Shopify's internal tooling. Tangle is described as a reproducible workflow system for ML and data experimentation. Tangent builds on that with auto-research loops that keep iterating toward better results across pipelines, prompts, throughput, and infrastructure — a capability Parakhin says is democratizing ML experimentation beyond dedicated researchers. SimGym simulates customers using Shopify's historical merchant data and browser-based agents, with Parakhin noting that this kind of simulation is difficult to replicate without years of real merchant behavior. The episode also touches on Liquid AI for ultra-low-latency and long-context workloads, the UCP product catalog as runtime infrastructure for agents, and Shopify's approach to PR review and CI/CD in an era where code generation has outpaced the rest of the software delivery pipeline.

Key facts

1. 01Shopify CTO Mikhail Parakhin appeared on the Latent Space podcast hosted by swyx.
2. 02A December 2025 model quality inflection point drove near-universal daily AI tool usage at Shopify.
3. 03Parakhin identifies PR review, CI/CD, and deployment stability — not code generation — as the real bottleneck in AI coding.
4. 04He warns that more AI-written code can still mean more bugs in production.
5. 05Running too many parallel agents that don't communicate is described as an anti-pattern that burns tokens inefficiently.
6. 06A critique loop where one agent's output is reviewed by a second, ideally different, model is said to produce significantly higher-quality code.
7. 07Shopify's internal AI stack includes Tangle (reproducible ML/data workflows), Tangent (auto-research loops), SimGym (customer simulation), and Liquid AI (low-latency and long-context workloads).