Shopify’s AI Phase Transition: 2026 Usage Explosion, Unlimited Opus-4.6 Token Budget, Tangle, Tangent, SimGym — with Mikhail Parakhin, Shopify CTO

Shopify has reached near-universal adoption of AI tools among its workforce, with a significant inflection point occurring in December 2025. Data indicates that daily active usage of AI tools now approaches 100% of the company. There is a clear shift toward CLI-based tools and internal agents that do not require direct code interaction, while traditional IDE-based tools are experiencing slower growth. The company funds unlimited tokens for employees, focusing on bottom-up model control rather than top-down restrictions, while encouraging the use of high-performance models to maintain quality.

The surge in AI-generated code necessitates rigorous automated PR reviews to prevent a spike in production bugs. Effective agentic workflows involve multiple agents—often using different models—critiquing and refining code rather than running numerous parallel agents that lack communication. The current CI/CD paradigm, designed for human-speed development, is becoming a bottleneck as machine-speed code generation increases the frequency of test failures and deployment cycles. Future development requires new metaphors for repository interaction and potentially a return to microservices to allow for independent, atomic deployments.

Tangle serves as a third-generation system for data processing and ML experimentation, utilizing content-based hashing to ensure efficiency and reproducibility. By automatically detecting when an experiment or data-processing task has already been performed, the system eliminates redundant work and prevents "digital archaeology." Tangent, an auto-research loop built on Tangle, allows agents to iterate on experiments, modify components, and optimize toward specific loss functions. This approach democratizes ML development, enabling product managers and non-ML engineers to perform complex research tasks without manual coding.

SimGym leverages decades of historical e-commerce data to create agents that replicate specific customer distributions and behaviors. By running simulations in headless browser environments, the system can perform counterfactual analysis—testing how interventions like discounts or layout changes affect conversion rates before they are implemented. This capability moves beyond traditional A/B testing by modeling complex, time-dependent user trajectories. The system requires significant infrastructure to run large-scale simulations and multi-modal models, but it provides a powerful mechanism for merchants to optimize their stores based on predicted outcomes.

Scaling AI operations requires specialized infrastructure to handle workloads that violate standard LLM serving assumptions, such as the need for multi-instance GPU configurations. Collaborations with partners like CentML and NVIDIA have been critical in optimizing models for specific profiles, whether prioritizing throughput, latency, or cost. Recent developments include bringing the entire Shopify product catalog into a unified search and identity-linking system. This allows for dynamic, real-time product lookups and personalized search results, providing a window into the vast universe of products sold across the platform.

Liquid neural networks offer a competitive non-transformer architecture that is particularly effective for low-latency and long-context applications. These models are sub-quadratic in context length and provide a compact, efficient way to represent data. They are increasingly used within Shopify for tasks ranging from 30-millisecond search query understanding to large-scale offline product categorization. While not a replacement for frontier models, these architectures serve as an excellent target for distillation, allowing for high-performance execution in batch jobs and latency-sensitive environments where traditional transformers may be too resource-intensive.

The development of AI assistants like Sydney demonstrated the importance of personality shaping, proving that a balance of politeness and "edginess" can significantly increase user engagement. This approach to personality design remains a key consideration for current projects like Sidekick. Shopify is actively seeking talent in machine learning, data science, and distributed systems, particularly those interested in reimagining database architecture through the lens of LLMs. The company continues to focus on building a platform that provides a network effect, enabling merchants to leverage advanced AI capabilities that would be impossible to implement at an individual scale.